ms-ai-architect/scripts/kb-eval/data/r71-payloads.json

227 lines
872 KiB
JSON
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

[
{
"file": "skills/ms-ai-engineering/references/agent-orchestration/foundry-workflows-visual-orchestration.md",
"claim_count": 16,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/agent-orchestration/foundry-workflows-visual-orchestration.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/agent-orchestration/foundry-workflows-visual-orchestration.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#1\",\n \"claim\": \"Foundry Workflows: Status Public Preview, annonsert på Microsoft Ignite november 2025\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/workflow\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#2\",\n \"claim\": \"Workflows er ett av tre agenttyper i Foundry: Prompt-based | Workflow | Hosted (preview, containeriserte agenter)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/workflow\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#3\",\n \"claim\": \"Node-typer i Foundry Workflows: Invoke agent | If/else | Go to | For each | Set variable | Parse value | Send message | Ask a question\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/workflow\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#4\",\n \"claim\": \"Foundry tilbyr tre startmaler for workflows: Sequential | Human-in-the-loop | Group chat\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/workflow\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#5\",\n \"claim\": \"Agent Framework HITL: RequestPort (C#) / ctx.request_info() (Python) for custom workflows; ToolApprovalRequestContent for agent orchestrations; workflow pauser og emitter RequestInfoEvent; checkpoints bevarer pending requests\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/agent-framework/workflows/human-in-the-loop\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#6\",\n \"claim\": \"Verktøy tilgjengelige i agent-noder i workflows (samme katalog som enkelt-agenter): Code Interpreter, Bing Search, Azure AI Search, Key Vault, MCP-servere\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/workflow\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#7\",\n \"claim\": \"Foundry MCP Server er i preview og eksponerer Foundry selv som MCP-endepunkt\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#8\",\n \"claim\": \"Azure Logic Apps: 1400+ triggere (HTTP, Events, Schedule, Queues, SaaS) | 400+ enterprise-koblinger (SAP, Salesforce, AS2, EDI)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#9\",\n \"claim\": \"Foundry Workflows datasuverenitet: Azure-regioner, Norway East støttet\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#10\",\n \"claim\": \"Agent Application-endepunkt kalles med api-version=2025-11-15-preview (protocols/openai/responses)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/how-to/agent-applications?view=foundry\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#11\",\n \"claim\": \"Klienter som kaller publisert Agent Application-endepunkt må ha rollen Foundry User (tidligere Azure AI User — rolle-ID og kjernepermisjoner uendret)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/how-to/agent-applications?view=foundry\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#12\",\n \"claim\": \"Per februar 2026: A/B-testing av workflows begrenset — all trafikk rutes til én aktiv deployment; trafikkdeling mellom deployments er på roadmap\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/how-to/agent-applications?view=foundry\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#13\",\n \"claim\": \"Foundry Workflows har immutable versjonssystem: hvert lagring oppretter ny versjon, eksisterende versjoner kan ikke endres; rollback via republisering; kodereferanse <agent_name>:<version>\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/development-lifecycle?view=foundry\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#14\",\n \"claim\": \"Agent Monitoring Dashboard-metrikker: Token usage | Latency | Run success rate | Error rate | Evaluation metrics\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/observability/how-to/how-to-monitor-agents-dashboard?view=foundry\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#15\",\n \"claim\": \"Foundry OpenTelemetry span-typer for multi-agent observability: execute_task | agent_to_agent_interaction | agent.state.management | agent_planning | execute_tool\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/foundry-workflows-visual-orchestration.md#16\",\n \"claim\": \"Declarative Workflows (Agent Framework) YAML action-typer: Variable, Control Flow, Output, Agent/Tool Invocation (InvokeAzureAgent, InvokeFunctionTool, InvokeMcpTool), HTTP, HITL (Question, RequestExternalInput), Conversation — C# og Python; Conversation-actions er C#-only\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/agent-framework/workflows/declarative\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/agent-orchestration/foundry-workflows-visual-orchestration.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/agent-orchestration/semantic-kernel-agents-implementation.md",
"claim_count": 8,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/agent-orchestration/semantic-kernel-agents-implementation.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/agent-orchestration/semantic-kernel-agents-implementation.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#1\",\n \"claim\": \"Microsoft Agent Framework 1.0 nådde GA 3. april 2026; produksjonsklart open-source rammeverk (.NET + Python)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/agent-framework/overview/agent-framework-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#2\",\n \"claim\": \"Semantic Kernel og AutoGen har konvergert inn i MAF og er nå i vedlikeholds-/migrasjonsmodus (migration guides finnes); ny utvikling bør bygge på MAF\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/agent-framework/overview/agent-framework-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#3\",\n \"claim\": \"De fem orkestreringsmønstrene Sequential, Concurrent, Handoff, Group Chat og Magentic er stabile (GA) i MAF 1.0 med streaming, checkpointing, human-in-the-loop og pause/resume\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/semantic-kernel/frameworks/agent/agent-orchestration/\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#4\",\n \"claim\": \"Semantic Kernel Agent Framework har fem hovedagent-typer: ChatCompletionAgent | OpenAIAssistantAgent | AzureAIAgent | OpenAIResponsesAgent | CopilotStudioAgent\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/semantic-kernel/frameworks/agent/agent-architecture\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#5\",\n \"claim\": \"Nøkkelforskjell: Semantic Kernel bruker Kernel-objektet som sentral orkestreringsenhet; Microsoft Agent Framework bruker IChatClient (Microsoft.Extensions.AI) wrapped av ChatClientAgent\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/agent-framework/overview/agent-framework-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#6\",\n \"claim\": \"MAF legger til enterprise-features: OpenTelemetry observability, Microsoft Entra-integrasjon, Agent-to-Agent (A2A) protocol, Model Context Protocol (MCP)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/agent-framework/overview/agent-framework-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#7\",\n \"claim\": \"Stateful agents (AzureAIAgent, OpenAIAssistantAgent) lagrer state serverside og krever matching thread-type (f.eks. AzureAIAgent + AzureAIAgentThread); ChatCompletionAgent er stateless med lokal ChatHistory\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/semantic-kernel/frameworks/agent/agent-architecture\"\n },\n {\n \"id\": \"ms-ai-engineering/agent-orchestration/semantic-kernel-agents-implementation.md#8\",\n \"claim\": \"Semantic Kernel: MIT License | Microsoft Agent Framework: MIT License\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/agent-orchestration/semantic-kernel-agents-implementation.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/api-management/apim-ai-gateway-overview.md",
"claim_count": 7,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/api-management/apim-ai-gateway-overview.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/api-management/apim-ai-gateway-overview.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#1\",\n \"claim\": \"APIM tier-støtte for AI gateway: Consumption = begrenset (ingen circuit breaker, semantic caching ja, ingen token limit by key); Developer/Basic v2/Standard v2 = full; Premium = full + multi-region\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#2\",\n \"claim\": \"Token-policies i APIM: azure-openai-token-limit (Azure OpenAI) | llm-token-limit (alle LLM-er) | azure-openai-emit-token-metric | llm-emit-token-metric\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#3\",\n \"claim\": \"APIM backend pool load balancing-moduser: Round-robin | Weighted | Priority-based | Session-aware (sesjonsaffinitet via cookie)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#4\",\n \"claim\": \"APIM circuit breaker støtter dynamisk trip duration basert på Retry-After header fra backend-respons\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#5\",\n \"claim\": \"Semantic caching i APIM krever Azure Managed Redis med RediSearch-modul\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#6\",\n \"claim\": \"APIM AI-sikkerhetsfunksjoner: llm-content-safety policy (automatisk moderering via Azure AI Content Safety) | Credential Manager for OAuth-autorisasjon | sikring av MCP-servere via APIM\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-ai-gateway-overview.md#7\",\n \"claim\": \"Bicep-ressurstype for APIM: Microsoft.ApiManagement/service@2023-09-01-preview (også backends; rolletildeling Microsoft.Authorization/roleAssignments@2022-04-01)\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/api-management/apim-ai-gateway-overview.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/api-management/apim-authentication-oauth-managed-identity.md",
"claim_count": 6,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/api-management/apim-authentication-oauth-managed-identity.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/api-management/apim-authentication-oauth-managed-identity.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/api-management/apim-authentication-oauth-managed-identity.md#1\",\n \"claim\": \"Microsoft Entra ID (tidligere Azure AD) er primær identitetsleverandør for Azure-tjenester\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-authentication-oauth-managed-identity.md#2\",\n \"claim\": \"RBAC-roller for Azure OpenAI: Cognitive Services OpenAI User (bruke deployments) | Cognitive Services OpenAI Contributor (opprette/administrere deployments) | Cognitive Services Contributor (full tilgang) | Reader (lese)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/api-management-authenticate-authorize-ai-apis\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-authentication-oauth-managed-identity.md#3\",\n \"claim\": \"APIM-policy validate-azure-ad-token finnes med attributter tenant-id, header-name, failed-validation-httpcode og elementer client-application-ids, audiences, required-claims\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/validate-azure-ad-token-policy\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-authentication-oauth-managed-identity.md#4\",\n \"claim\": \"APIM-policy authentication-managed-identity finnes med attributter resource, output-token-variable-name, ignore-error for tokenhenting via managed identity\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/authentication-managed-identity-policy\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-authentication-oauth-managed-identity.md#5\",\n \"claim\": \"Ved import av API direkte fra Microsoft Foundry konfigurerer APIM automatisk managed identity-autentisering mot backend\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/azure-openai-api-from-specification\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/apim-authentication-oauth-managed-identity.md#6\",\n \"claim\": \"I APIM v2-tiers kan custom CA-sertifikater konfigureres direkte på backend-entiteten (thumbprint, subject name + issuer thumbprint, chain validation)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/api-management/apim-authentication-oauth-managed-identity.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/api-management/backend-pool-management.md",
"claim_count": 10,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/api-management/backend-pool-management.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/api-management/backend-pool-management.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#1\",\n \"claim\": \"APIM backend pool-typer for AI-workloads: Round-robin | Weighted | Priority-based | Session-aware (sticky sessions)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#2\",\n \"claim\": \"Session-aware pools konfigureres med sessionAffinity (sessionId source: Cookie); session awareness bruker Set-Cookie header og klienten må håndtere cookies\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#3\",\n \"claim\": \"APIM bruker circuit breaker-regler for å vurdere backend-helse i stedet for tradisjonelle health probes; ved trip rutes trafikk til neste prioritetsgruppe\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#4\",\n \"claim\": \"Circuit breaker-egenskapen acceptRetryAfter: true gjør at breakeren respekterer Retry-After header fra backend (bl.a. Azure OpenAI 429)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#5\",\n \"claim\": \"Azure OpenAI returnerer Retry-After header ved 429-responser; verdien kan være opptil 1 dag ved alvorlig overbelastning\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#6\",\n \"claim\": \"APIM innebygd status-endepunkt: GET https://{apim-name}-{region}-01.regional.azure-api.net/status-0123456789abcdef\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#7\",\n \"claim\": \"APIM-begrensning: maks 30 backends per pool\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#8\",\n \"claim\": \"Bicep-ressurstype for APIM-backends: Microsoft.ApiManagement/service/backends@2023-09-01-preview\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#9\",\n \"claim\": \"Circuit breaker-regler i APIM er tilnærmede (approximate) pga. distribuert arkitektur — ulike gateway-instanser synkroniserer ikke circuit state\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/api-management/backend-pool-management.md#10\",\n \"claim\": \"Log Analytics-tabellen ApiManagementGatewayLlmLog finnes med kolonner TotalTokens, PromptTokens, CompletionTokens, BackendUrl (og ApiManagementGatewayLogs med BackendResponseCode)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/api-management/backend-pool-management.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-api-best-practices.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-api-best-practices.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-api-best-practices.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#1\",\n \"claim\": \"Azure OpenAI SDK-er (Python, .NET, Go) retrier automatisk 408|429|500|502|503|504 — opptil 3 ganger med eksponentiell backoff\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#2\",\n \"claim\": \"OpenAI Python SDK (Azure): api_version=2024-10-21|max_retries default: 2\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#3\",\n \"claim\": \"Azure OpenAI Batch API-limits: max batch files uten utløp 500|med utløp 10 000|max input file size 200 MB (BYOS: 1 GB)|max requests per fil 100 000\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#4\",\n \"claim\": \"Batch API: asynkrone batch-operasjoner med 50% lavere kostnad enn real-time API\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#5\",\n \"claim\": \"Fail-fast-regioner for batching: Australia East|East US|Germany West Central|Italy North|North Central US|Poland Central|Sweden Central|Switzerland North|East US 2|West US\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#6\",\n \"claim\": \"Azure OpenAI Python SDK default timeout: 600s (10 min)\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#7\",\n \"claim\": \"Azure Service Bus duplicate detection-vindu: default 10 minutter|min 20 sekunder|max 7 dager; basert på MessageId (evt. + PartitionKey)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/service-bus-messaging/duplicate-detection\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#8\",\n \"claim\": \"Ved 429 på provisioned deployments returnerer tjenesten retry-after og retry-after-ms headers; standard SDK-oppførsel respekterer retry-after\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/provisioned-get-started\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#9\",\n \"claim\": \"Azure Monitor-metrikk for PTU-utnyttelse heter Provisioned-Managed Utilization V2\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/provisioned-get-started\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#10\",\n \"claim\": \"Provisioned reservation-rabatter: 1-årig commitment ~37%|3-årig commitment ~57%\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/provisioned-get-started\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#11\",\n \"claim\": \"SDK-er som støtter Microsoft Foundry: Python azure-ai-inference + openai (Azure-variant)|.NET Azure.AI.Inference + Azure.AI.OpenAI|JS/TS @azure/openai + @azure/ai-inference|Go github.com/openai/openai-go (med Azure endpoint)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-api-best-practices.md#12\",\n \"claim\": \"429-feil koster ingenting (ingen PTU-forbruk), men 400 content filter-feil, 408 timeout og finish_reason: content_filter koster\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/provisioned-get-started\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-api-best-practices.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-networking-security.md",
"claim_count": 15,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-networking-security.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-networking-security.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#1\",\n \"claim\": \"Microsoft har omdøpt Azure Cognitive Services til Microsoft Foundry Tools (2026-04); API-endepunkter, RBAC-roller og ARM-ressurstyper beholder CognitiveServices i identifikatoren\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#2\",\n \"claim\": \"Private DNS-soner: privatelink.cognitiveservices.azure.com (AI Services)|privatelink.openai.azure.com (Azure OpenAI)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#3\",\n \"claim\": \"Target sub-resource for private endpoint (AI Services, Azure OpenAI): account\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#4\",\n \"claim\": \"Service tag for service endpoints: Microsoft.CognitiveServices\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#5\",\n \"claim\": \"Maks 100 VNet-regler per ressurs (kan kombineres med IP-regler)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#6\",\n \"claim\": \"IP firewall-regler: kun offentlige IPv4-adresser; private RFC 1918-områder (10.*, 172.16-31.*, 192.168.*) ikke tillatt\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#7\",\n \"claim\": \"Azure AI Services støtter ikke direkte VNet-injeksjon (deployes ikke inn i kundens VNet); bruker private endpoints i stedet\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#8\",\n \"claim\": \"Nettverkssikkerhet (private endpoints, IP firewall, service endpoints) krever Basic tier (S0) eller høyere; Free tier ikke støttet\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#9\",\n \"claim\": \"RBAC-roller for tilgang: Cognitive Services User|Cognitive Services OpenAI User\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#10\",\n \"claim\": \"Trusted services bypass aktiveres med networkAcls.bypass: AzureServices; tillater Azure AI Search og Azure Machine Learning å kalle tjenesten via managed identity (JWT-validering)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/on-your-data-configuration\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#11\",\n \"claim\": \"Custom DNS (on-premises) krever conditional forwarder til Azure DNS 168.63.129.16\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#12\",\n \"claim\": \"Service tags for portal-tilgang: AzureActiveDirectory|AzureFrontDoor.Frontend|AzureResourceManager|CognitiveServicesManagement|CognitiveServicesFrontEnd\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#13\",\n \"claim\": \"NSG-støtte ikke tilgjengelig for Azure AI Services; private link og disable public access støttet (per security baseline)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/security/benchmark/azure/baselines/cognitive-services-security-baseline\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#14\",\n \"claim\": \"Custom subdomain er påkrevd for private endpoints på AI Services\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-networking-security.md#15\",\n \"claim\": \"Power Automate/Power Apps mot private endpoints krever On-premises data gateway; Azure Relay hybrid connection ikke støttet direkte; custom connector må bruke public endpoint med IP firewall\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-networking-security.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md",
"claim_count": 16,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#1\",\n \"claim\": \"Microsoft har endret terminologi Cognitive Services → Foundry Tools (2025); resource type kind: AIServices er konsistent\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#2\",\n \"claim\": \"Personalizer er utgått (retired)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/architecture/data-guide/technology-choices/ai-services\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#3\",\n \"claim\": \"Foundry resource type: kind AIServices med capabilities Agents|Evaluations|Azure OpenAI|Speech|Vision|Language|Content Understanding; legacy Azure OpenAI resource: kind OpenAI (kun OpenAI-modeller)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/concepts/resource-types\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#4\",\n \"claim\": \"Azure AI Hub (kind: Hub) er deprecated og migreres til Foundry resource type (Juni 2025+)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/concepts/resource-types\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#5\",\n \"claim\": \"Foundry Model Catalog: over 1 900 modeller fra Microsoft, OpenAI, Anthropic, Meta, Mistral, Cohere\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#6\",\n \"claim\": \"Azure OpenAI modellserie 2026-06: GPT-5-generasjonen (GPT-5/5.1/5.2/5.4/5.5 + -mini/-nano/-codex) er flaggskip for reasoning/chat; GPT-5.5 nyest\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#7\",\n \"claim\": \"GPT-4.1, GPT-4.1-mini, GPT-4.1-nano: ~1M tokens kontekst; deployment-typer Standard|Global Standard|Provisioned\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#8\",\n \"claim\": \"o4-mini, o3, o3-mini er eldre o-serie reasoning-modeller, erstattet av GPT-5-generasjonen for nye design; deployment Standard|Provisioned\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#9\",\n \"claim\": \"Global Standard deployment: globalt routet, ingen data residency, 9 % rimeligere, høy throughput\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#10\",\n \"claim\": \"Azure OpenAI er inkludert i Foundry resource type (kind: AIServices); legacy Azure OpenAI-ressurser (kind: OpenAI) kan oppgraderes til Foundry uten API-endringer\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/concepts/resource-types\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#11\",\n \"claim\": \"Content Understanding: GA v1.0 / 2025-11-01; multimodal, generative fields, reasoning\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/choosing-right-ai-tool\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#12\",\n \"claim\": \"QnA Maker er utgått (retired); erstattes av Language Understanding for FAQ-bots\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#13\",\n \"claim\": \"Norge-spesifikke Azure-regioner: Norway East (Oslo)|Norway West (Stavanger)\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#14\",\n \"claim\": \"On-prem deployment: Foundry Tools ja (containers)|Microsoft Foundry nei (cloud-only)|Azure OpenAI nei (cloud-only)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#15\",\n \"claim\": \"Commitment Tiers (Foundry Tools): fast månedskostnad, spar 30-40 % vs pay-per-use\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/concepts/manage-costs\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#16\",\n \"claim\": \"Agent-publisering til Agent 365 krever M365 E3 eller høyere (Azure subscription + M365 E3/E5)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/mlops-genaiops/prompt-flow-production-deployment.md",
"claim_count": 13,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/mlops-genaiops/prompt-flow-production-deployment.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/mlops-genaiops/prompt-flow-production-deployment.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#1\",\n \"claim\": \"Prompt Flow (både Microsoft Foundry og Azure ML) pensjoneres 2027-04-20; anbefales ikke for ny utvikling; migrer til Microsoft Agent Framework; web authoring, VS Code-utvidelser og container images (promptflow-runtime, promptflow-runtime-stable, promptflow-python) får ikke lenger oppdateringer inkl. sikkerhetsoppdateringer\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/machine-learning/prompt-flow/migrate-prompt-flow-to-agent-framework?view=azureml-api-2\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#2\",\n \"claim\": \"Default base image: mcr.microsoft.com/azureml/promptflow/promptflow-runtime-stable:latest\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-deploy-to-code?view=azureml-api-2\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#3\",\n \"claim\": \"PROMPTFLOW_SERVING_ENGINE: flask (default)|fastapi krever SDK >= 1.10.0\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-deploy-to-code?view=azureml-api-2\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#4\",\n \"claim\": \"PROMPTFLOW_WORKER_NUM default = CPU cores|PROMPTFLOW_WORKER_THREADS default = 1\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-deploy-to-code?view=azureml-api-2\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#5\",\n \"claim\": \"Azure Managed Online Endpoint krever Microsoft.PolicyInsights resource provider registrert\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/flow-deploy?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#6\",\n \"claim\": \"Deployment-identitet krever rollen Azure Machine Learning Workspace Connection Secrets Reader; connections med Entra ID auth (Azure OpenAI) krever Cognitive Services OpenAI User\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/flow-deploy?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#7\",\n \"claim\": \"Deployment tar 15-20 minutter (endpoint creation, model registration, deployment creation)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/flow-deploy?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#8\",\n \"claim\": \"App Insights-metrics under namespace 'promptflow standard metrics': token_consumption|flow_latency|flow_request|node_latency|node_request|rpc_latency|rpc_request|flow_streaming_response_duration\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/develop/trace-production-sdk?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#9\",\n \"claim\": \"Prompt Flow serving eksponerer /feedback endpoint for post-inference feedback; traceparent-header korrelerer feedback til original request trace\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/develop/trace-production-sdk?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#10\",\n \"claim\": \"Prompt Flow er open source (MIT license) + Azure-managed variant; ingen lisenskostnad for SDK/CLI\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#11\",\n \"claim\": \"Reserved instances: opptil 72% rabatt for forutsigbar last\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#12\",\n \"claim\": \"Concurrency-formel: max_concurrent_requests_per_instance = worker_num × worker_threads × multiplier (1.0 hvis request time > 200ms, 1.5-2.0 hvis <= 200ms)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-deploy-to-code?view=azureml-api-2\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/prompt-flow-production-deployment.md#13\",\n \"claim\": \"Instance count minimum 3 for high availability ved deployment\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/flow-deploy?view=foundry-classic\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/mlops-genaiops/prompt-flow-production-deployment.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/rag-architecture/citation-tracking.md",
"claim_count": 10,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/rag-architecture/citation-tracking.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/rag-architecture/citation-tracking.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#1\",\n \"claim\": \"Azure AI Search agentic retrieval: grounding-overflaten (references + activity-log) er GA via REST 2026-04-01; LLM query planning (ikke-minimal reasoning effort) + answer synthesis + multi-turn messages forblir preview (2026-05-01-preview); portal/Foundry preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#2\",\n \"claim\": \"Azure AI Search REST API-versjoner for agentic retrieval: 2026-04-01 (GA) og 2026-05-01-preview\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#3\",\n \"claim\": \"Classic RAG (Azure AI Search + Azure OpenAI, 'On Your Data' med citation annotations) er GA\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#4\",\n \"claim\": \"Semantic Ranking-score (Azure AI Search): skala 0.04.0\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#5\",\n \"claim\": \"Vector Similarity (Cosine) i Azure AI Search: skala 0.3331.0\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#6\",\n \"claim\": \"Groundedness Pro: threshold-basert scoring, leveres av Azure AI Content Safety (per evaluerings-kall)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#7\",\n \"claim\": \"Semantic Answer Confidence (Azure AI Search): 70% terskel\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#8\",\n \"claim\": \"Semantic Ranker prising: gratis månedlig request-kvote + standard pay-as-you-go etter kvoten\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#9\",\n \"claim\": \"Agentic retrieval faktureres token-basert (retrieval-tokens, ikke per query): månedlig gratis-kvote, deretter pay-as-you-go; Azure OpenAI faktureres separat for query planning/answer synthesis\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/citation-tracking.md#10\",\n \"claim\": \"Azure AI Content Safety tilbyr groundedness-deteksjon med korreksjonsfunksjon\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/rag-architecture/citation-tracking.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md",
"claim_count": 11,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#1\",\n \"claim\": \"GPT-4 Turbo (gpt-4 turbo-2024-04-09): 128k tokens context window\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#2\",\n \"claim\": \"GPT-4.1 series: ~1M tokens context window (max 1 047 576; standard deployments 300k; provisioned managed/batch 128k)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#3\",\n \"claim\": \"GPT-4o: 128k tokens context window | gpt-4o-mini: 128k tokens context window\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#4\",\n \"claim\": \"o1: 200k context window | o3-mini: 200k input / 100k output\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#5\",\n \"claim\": \"Azure OpenAI quota system is Quota Tiers: Free Tier (Tier 0) plus Tier 1-6 (Tier 6 highest), quota auto-increases with usage; replaces former Default/Enterprise two-tier system; TPM/RPM defined per region, per subscription, per model/deployment type\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#6\",\n \"claim\": \"GlobalStandard TPM examples (Tier 1 to Tier 6): gpt-4.1 1M to 45M | gpt-4o-mini 2M to 225M | o1 3M to 48M | o3-mini 5M to 80M; max TPM (Tier 6) = 225M\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#7\",\n \"claim\": \"Azure OpenAI Assistants API supports max_prompt_tokens and max_completion_tokens on Run creation; truncation_strategy options are auto and last_messages\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/concepts/assistants#context-window-management\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#8\",\n \"claim\": \"For Assistants API File Search tool: set max_prompt_tokens to minimum 20,000; for longer conversations 50,000 or remove limit\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/concepts/assistants#context-window-management\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#9\",\n \"claim\": \"Azure OpenAI Batch API gives 50% discount for non-interactive workloads\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#10\",\n \"claim\": \"Azure OpenAI Batch quota (per Quota Tier): gpt-4.1 ~500M tokens/month (upper tier), ~30M (lower tier); gpt-4o ~500M (upper), ~30M (lower)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-context-windows.md#11\",\n \"claim\": \"Prompt Flow (Microsoft Foundry) shows estimated token count per node in the flow\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-governance/references/responsible-ai/content-safety-implementation.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-governance/references/responsible-ai/content-safety-implementation.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-governance/references/responsible-ai/content-safety-implementation.md`)\n\n[\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#1\",\n \"claim\": \"Content Safety feature statuses: Analyze Text GA | Analyze Image GA | Prompt Shields GA | Groundedness Detection Preview | Protected Material Text GA | Protected Material Code GA | Custom Categories (Standard) Preview | Custom Categories (Rapid) Preview | Blocklists GA\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#2\",\n \"claim\": \"Content Safety input limits: Analyze Text max 10K chars | Analyze Image JPEG/PNG/GIF/BMP/TIFF/WEBP max 4MB | Prompt Shields text max 10K chars | Groundedness query+sources max 55K chars | Protected Material Text min 110 chars\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#3\",\n \"claim\": \"Content analysis uses four harm categories (hate, sexual, violence, self-harm) with severity levels 0-6 (0 Safe, 2 Low, 4 Medium, 6 High); 0 is annotate-only, 2/4/6 configurable\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/harm-categories\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#4\",\n \"claim\": \"Azure OpenAI default content filter blocks Medium (4) and High (6) severity for all four categories\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/default-safety-policies\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#5\",\n \"claim\": \"Content filter response behavior: prompt blocked gives HTTP 400 with error.code=content_filter | completion blocked (non-streaming) gives 200 with finish_reason=content_filter | streaming blocked: stream stops, last chunk carries finish_reason | filter error: content_filter_results.error populated\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/foundry-models/concepts/content-filter\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#6\",\n \"claim\": \"Copilot Studio cannot (per feb 2026) configure severity levels per category - uses Azure OpenAI deployment settings; custom blocklists can be enabled in Agent Settings\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#7\",\n \"claim\": \"AI Builder Text generation uses Azure OpenAI with content filtering enabled by default and no configuration options (default Medium+High block)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#8\",\n \"claim\": \"Microsoft 365 Copilot has its own content filtering policies that are not customer-configurable (severity levels cannot be adjusted; Microsoft-managed)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#9\",\n \"claim\": \"Azure AI Content Safety available in West Europe and Norway East (via Azure OpenAI); models run in EU (no data transfer to USA)\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#10\",\n \"claim\": \"Azure AI Content Safety has PII detection for completions (names, addresses; Norwegian national ID format not officially supported), configurable to block or mask PII in LLM output\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#11\",\n \"claim\": \"Customer-managed keys (CMK/BYOK) supported for Content Safety blocklist data; blocklist data stored in same region as resource, encrypted at rest; no prompts/completions stored for training\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/content-safety/data-privacy\"\n },\n {\n \"id\": \"ms-ai-governance/responsible-ai/content-safety-implementation.md#12\",\n \"claim\": \"Content Safety tiers/rate limits: F0 (Free) 5 RPS | S0 (Standard) 1000 RP10S | Groundedness 50 RPS limit\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-governance/references/responsible-ai/content-safety-implementation.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/ai-security-engineering/ai-prompt-shield-network.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/ai-security-engineering/ai-prompt-shield-network.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/ai-security-engineering/ai-prompt-shield-network.md`)\n\n[\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#1\",\n \"claim\": \"Two separate products: Azure AI Content Safety Prompt Shields is GA; AI Gateway Prompt Shield via Global Secure Access is Preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/global-secure-access/how-to-ai-prompt-shield\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#2\",\n \"claim\": \"Prompt Shields endpoint: POST {endpoint}/contentsafety/text:shieldPrompt?api-version=2024-09-01\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/content-safety/quickstart-jailbreak\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#3\",\n \"claim\": \"Prompt Shields input limits: userPrompt max 10 000 chars | documents array max 5 documents per request | each document max 10 000 chars\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/content-safety/concepts/jailbreak-detection\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#4\",\n \"claim\": \"AI Gateway Prompt Shield has pre-configured extractors for: Microsoft Copilot, OpenAI ChatGPT, Anthropic Claude, Meta Llama, xAI Grok, Mistral, Cohere, Inflection Pi, Alibaba Qwen, plus custom JSON-based LLMs (custom URL + JSON path)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/global-secure-access/how-to-ai-prompt-shield\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#5\",\n \"claim\": \"AI Gateway Prompt Shield limitations: text prompts only (no files); JSON-based GenAI apps only (not URL-encoded, e.g. Gemini); max 10 000 chars per prompt (longer prompts truncated)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/global-secure-access/how-to-ai-prompt-shield\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#6\",\n \"claim\": \"AI Gateway Prompt Shield prerequisites: Microsoft Entra Internet Access license, Entra-joined or hybrid-joined devices, Global Secure Access Administrator and Conditional Access Administrator roles\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/global-secure-access/how-to-ai-prompt-shield\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#7\",\n \"claim\": \"APIM AI gateway mediates OpenAI Chat Completions/Responses, Anthropic Messages API (supported in v2 tiers), Google Vertex AI, remote MCP servers and A2A Agent APIs; unified model API (preview) exposes multiple backends via one OpenAI-compatible endpoint; AI gateway can integrate directly with Microsoft Foundry (preview)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/genai-gateway-capabilities#security-and-safety\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#8\",\n \"claim\": \"APIM llm-content-safety policy: shield-prompt attribute; enforce-on-completions (inbound, also checks LLM responses); window-size default 10 000 chars; window-overlap-size; threshold 0-7; output-type FourSeverityLevels (0,2,4,6) or EightSeverityLevels (0-7); blocked requests return 403; supported categories Hate/SelfHarm/Sexual/Violence; supports blocklists; usable in inbound and outbound\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/llm-content-safety-policy\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#9\",\n \"claim\": \"APIM llm-content-safety policy available in tiers: Developer, Basic, Basic v2, Standard, Standard v2, Premium, Premium v2\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/api-management/llm-content-safety-policy\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#10\",\n \"claim\": \"Groundedness Detection: supports QnA and summarization tasks; domain options MEDICAL or GENERIC; two modes Non-Reasoning and Reasoning; Groundedness Correction is preview (returns correctedText); only English text supported\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#11\",\n \"claim\": \"Groundedness endpoint: POST {endpoint}/contentsafety/text:detectGroundedness?api-version=2024-09-01\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/ai-prompt-shield-network.md#12\",\n \"claim\": \"AI Gateway Prompt Shield (Preview) is included in the Microsoft Entra Internet Access license (licensed per user/month)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/global-secure-access/how-to-ai-prompt-shield\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/ai-security-engineering/ai-prompt-shield-network.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/ai-security-engineering/zero-trust-ai-services.md",
"claim_count": 11,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/ai-security-engineering/zero-trust-ai-services.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/ai-security-engineering/zero-trust-ai-services.md`)\n\n[\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#1\",\n \"claim\": \"RBAC roles: Cognitive Services OpenAI User (5e0bd9bd-7b93-4f28-af87-19fc36ad61bd, inference) | Cognitive Services OpenAI Contributor (a001fd3d-188f-4b5d-821b-7da978bf7442, inference + deployment) | Cognitive Services User (a97b65f3-24c7-4388-baec-2e87135dc908, data plane) | Cognitive Services Contributor (25fbc0a9-bd7c-42a3-aa1a-3b75d497ee68, full control incl. keys)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/document-translation/how-to-guides/create-use-managed-identities\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#2\",\n \"claim\": \"Managed Identity tokens are cached up to 24 hours; group/role membership changes can take hours to propagate; App Roles propagate faster than groups\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/entra/identity/managed-identities-azure-resources/managed-identity-best-practice-recommendations\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#3\",\n \"claim\": \"Microsoft now explicitly recommends avoiding API keys for Azure AI Services in production; Microsoft Foundry resource (formerly 'Foundry Tools', kind=AIServices) uses the same DefaultAzureCredential pattern across all AI services\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/managed-identity\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#4\",\n \"claim\": \"CAE token lifetime: up to 28 hours (vs standard access token 1 hour, Managed Identity refresh token 24 hours); critical events revoke in seconds\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-continuous-access-evaluation\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#5\",\n \"claim\": \"From 2025 CAE is configured via Conditional Access policies (Session controls) instead of a separate toggle; options Disabled, Basic (default), Strict\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-continuous-access-evaluation\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#6\",\n \"claim\": \"Strict Location Enforcement is Preview, activated per Conditional Access policy, blocks tokens used outside approved network locations in near-realtime, and now supports Azure Government clouds\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-continuous-access-evaluation\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#7\",\n \"claim\": \"AKS Workload Identity replaces AAD Pod Identity for giving pods Managed Identity\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#8\",\n \"claim\": \"Microsoft Foundry Managed Network Isolation modes: Allow Internet Outbound (default) and Allow Only Approved Outbound; Foundry auto-creates private endpoints\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/how-to-network-isolation-planning\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#9\",\n \"claim\": \"Virtual Network Integration for Dataverse is in Private Preview (Q1 2026)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/power-platform/admin/vnet-data-gateway\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#10\",\n \"claim\": \"Zero Trust Guidance Center is organized by technology pillars and business scenarios and now includes a dedicated 'Protect AI and data' scenario (security-adoption-scenario-secure-ai); Secure Future Initiative (SFI) has its own 'SFI patterns and best practices' section in the Zero Trust hub\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/security/zero-trust/\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/zero-trust-ai-services.md#11\",\n \"claim\": \"Azure AI security baseline (azure-openai-security-baseline) reflects Microsoft Cloud Security Benchmark (MCSB) v1.0\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/security/benchmark/azure/baselines/azure-openai-security-baseline\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/ai-security-engineering/zero-trust-ai-services.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md",
"claim_count": 9,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md`)\n\n[\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#1\",\n \"claim\": \"Model Quota (TPM) scope: per subscription, per region, per model; default varies per Quota Tier (Free Tier (Tier 0) + Tier 1-6); ca. 100K-225M TPM; adjustable via quota request\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#2\",\n \"claim\": \"VM Family Quota default 24-300 cores per subscription per region (depends on subscription type), adjustable via support request\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/hub-quota\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#3\",\n \"claim\": \"Compute Instances: 500 total compute limit per region, adjustable up to 2500 via quota UI, beyond via support\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/hub-quota\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#4\",\n \"claim\": \"Serverless API Quota: 200K TPM and 1K RPM per deployment; one deployment per model per project by default\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#5\",\n \"claim\": \"Azure OpenAI has no hard-limit budget enforcement (unlike OpenAI API); Azure budgets only send alerts and do not stop consumption automatically\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/concepts/manage-costs\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#6\",\n \"claim\": \"Dynamic quota is Preview; lets deployments opportunistically use unused capacity beyond baseline quota; has no upper cap for spending control\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/dynamic-quota\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#7\",\n \"claim\": \"The underlying Azure resource (API kind AIServices) is renamed Microsoft Foundry resource (successor/renaming of 'Foundry Tools'); projects function as folders grouping work under one resource\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#8\",\n \"claim\": \"Foundry RBAC roles renamed (role IDs and core permissions unchanged): Azure AI User to Foundry User | Azure AI Owner to Foundry Owner | Azure AI Account Owner to Foundry Account Owner | Azure AI Project Manager to Foundry Project Manager\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#9\",\n \"claim\": \"Azure Owner and Contributor give only management access (create/configure resources, quota/cost) - not development data actions; Foundry User (least privilege for developers) or Foundry Owner required to build\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/cost-optimization/batch-processing-cost-reduction.md",
"claim_count": 11,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/cost-optimization/batch-processing-cost-reduction.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/cost-optimization/batch-processing-cost-reduction.md`)\n\n[\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#1\",\n \"claim\": \"Azure OpenAI Batch API gir 50% kostnadsreduksjon vs standard global deployment og bruker separat enqueued token-kvote\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#2\",\n \"claim\": \"To batch deployment-typer: Global-Batch (globalt distribuert) | Data Zone Batch (regionsbasert)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#3\",\n \"claim\": \"Exponential backoff (automatisk retry av store batch-jobber) støttes i utvalgte regioner\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#4\",\n \"claim\": \"Batch-jobber har 24-timers target completion window; jobber som tar lengre tid expires ikke; kansellering når som helst, betaler kun for fullført arbeid\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#5\",\n \"claim\": \"Batch-støttede modeller (juni 2026): gpt-5.1 2025-11-13 | gpt-5 2025-08-07 | o3 2025-04-16 | o4-mini 2025-04-16 | gpt-4.1 2025-04-14 | gpt-4.1-mini 2025-04-14 | gpt-4.1-nano 2025-04-14 | o3-mini 2025-01-31 | gpt-4o 2024-05-13/2024-08-06/2024-11-20 | gpt-4o-mini 2024-07-18\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#6\",\n \"claim\": \"Tilgang til gpt-5 og o3 i batch krever registrering\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#7\",\n \"claim\": \"gpt-5.1, gpt-5 og o3-mini er ikke tilgjengelige i alle Data Zone Batch-regioner\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#8\",\n \"claim\": \"o3-mini er ikke støttet med API-versjon 2024-10-21; nyere modeller med Global Batch krever v1 API\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#9\",\n \"claim\": \"Batch støttes ikke med Assistants API eller Azure OpenAI On Your Data\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/how-to/batch\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#10\",\n \"claim\": \"Batch input file (JSONL, purpose \\\"batch\\\") kan settes med expiration 14-30 dager fra upload\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/batch?pivots=programming-language-python#upload-batch-file\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/batch-processing-cost-reduction.md#11\",\n \"claim\": \"Azure OpenAI Batch API har status GA\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/whats-new#august-2024\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/cost-optimization/batch-processing-cost-reduction.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/cost-optimization/token-counting-optimization.md",
"claim_count": 16,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/cost-optimization/token-counting-optimization.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/cost-optimization/token-counting-optimization.md`)\n\n[\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#1\",\n \"claim\": \"o200k_base er default tiktoken-encoding for gpt-4o, o1, o3\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://github.com/openai/tiktoken\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#2\",\n \"claim\": \"Azure OpenAI On Your Data er deprecated og pensjoneres 2026-10-14; migrer til Foundry Agent Service + Foundry IQ\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/concepts/use-your-data#token-usage-estimation-for-azure-openai-on-your-data\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#3\",\n \"claim\": \"On Your Data RAG token-fordeling: 20% av context window reservert for model response, 80% deles mellom meta prompt/spørsmål/historikk/chunks; user question og history capped ved 2000 tokens\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/concepts/use-your-data#token-usage-estimation-for-azure-openai-on-your-data\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#4\",\n \"claim\": \"Prompt caching automatisk aktivert for GPT-4o og nyere; minimum prompt length 1024 tokens for cache hit; granularitet 128 tokens etter første 1024; in-memory cache TTL 5-10 min idle, max 1 time\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#5\",\n \"claim\": \"Extended cache TTL opptil 24 timer via prompt_cache_retention:\\\"24h\\\" for GPT-5-serien + gpt-4.1; default 24h for nyere modeller\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#6\",\n \"claim\": \"Cached tokens: 50% rabatt på Standard deployment; opptil 100% rabatt på Provisioned (inkludert i PTU-pris)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#7\",\n \"claim\": \"Samme prefix + prompt_cache_key over ~15 req/min gjør at noen requests overflower til ekstra maskiner og cache-effektiviteten faller\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#8\",\n \"claim\": \"Token-limit per chunk for embeddings: text-embedding-ada-002 8191 tokens | text-embedding-3-small/large 8191 tokens\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#9\",\n \"claim\": \"gpt-4o-mini fine-tuning: training example max 64536 tokens, input limit 128000 tokens\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#10\",\n \"claim\": \"On Your Data gjennomsnittlig token-bruk per request: intent prompt ~1366 | generation prompt ~4297 | response ~111 | intent output ~25 | totalt ~5800 tokens\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/concepts/use-your-data#token-usage-estimation-for-azure-openai-on-your-data\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#11\",\n \"claim\": \"RFT (Reinforcement Fine-Tuning) faktureres tidsbasert (ikke token-basert) med per-job-tak $5000; ved tak pauses training og deploybart checkpoint opprettes; ingen kostnad for kø/failed jobs/kansellering før training\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning-cost-management\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#12\",\n \"claim\": \"Fine-tuned model hosting: $1.70/time per deployment (Standard/Global Standard), påløper selv uten bruk\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning-cost-management\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#13\",\n \"claim\": \"Developer tier fine-tuning: ingen hosting fee, men deployment auto-deletes etter 24 timer\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/fine-tuning-cost-management\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#14\",\n \"claim\": \"Context windows: gpt-5 400K | gpt-5-mini 400K | gpt-4o 128K | gpt-4o-mini 128K | o3-mini 200K | gpt-4.1 128K\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#15\",\n \"claim\": \"Image token conversion: low-res (<512x512) 85 tokens flat; high-res resizes til 2048x2048, splittes i 512x512 tiles, 170 tokens per tile + 85 base\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/token-counting-optimization.md#16\",\n \"claim\": \"Copilot Studio: token-basert billing for Generative Answers (Azure OpenAI), message-basert billing for standard topics; token counting via AI Builder credits\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/cost-optimization/token-counting-optimization.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/cost-optimization/gpt5-gpt41-pricing-models.md",
"claim_count": 17,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/cost-optimization/gpt5-gpt41-pricing-models.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/cost-optimization/gpt5-gpt41-pricing-models.md`)\n\n[\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#1\",\n \"claim\": \"gpt-4.1 kontekstvindu: 1047576 tokens (full), 128000 tokens (standard og provisioned deployments), 300000 tokens (batch deployments)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#2\",\n \"claim\": \"gpt-4.1: treningsdata til og med mai 2024; versjon 2025-04-14\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#3\",\n \"claim\": \"GPT-4.1-serien deployment-typer: Global Standard, Data Zone Standard, Regional (Standard og Provisioned)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#4\",\n \"claim\": \"Priority Processing er tilgjengelig for gpt-4.1, ikke for gpt-4.1-mini/nano\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#5\",\n \"claim\": \"gpt-5 og gpt-5-codex krever registrering og godkjenning; gpt-5-mini, gpt-5-nano, gpt-5-chat har ingen registreringskrav\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#6\",\n \"claim\": \"GPT-5 kontekstvindu: 400000 tokens (272K input / 128K output for resonneringsmodeller)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#7\",\n \"claim\": \"PTU-gjennomstrømming (input TPM per PTU / min PTU global / min PTU regional): gpt-5 4750/15/50 | gpt-5-mini 23750/15/25 | gpt-4.1 3000/15/50 | gpt-4.1-mini 14900/15/25 | gpt-4.1-nano 59400/15/25 | o4-mini 5400/15/25; latens-SLA p50: 99% > 50/80/80/90/100/90 TPS\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/concepts/provisioned-throughput-billing?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#8\",\n \"claim\": \"PTU input/output-ratio: gpt-4.1 1 output token = 4 input tokens; gpt-5 1 output token = 8 input tokens\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/openai/concepts/provisioned-throughput-billing?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#9\",\n \"claim\": \"Copilot Credits-takstnivåer: gpt-4.1-mini Basic | gpt-4.1 Standard | gpt-5-chat (preview) Standard | gpt-5-reasoning (preview) Premium | o3 Premium | Claude Sonnet 4.5 (experimental) Standard | Claude Opus 4.5 (experimental) Premium\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/microsoft-copilot-studio/prompt-model-settings\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#10\",\n \"claim\": \"gpt-4.1-mini er standardmodellen i Copilot Studio (erstattet GPT-4o mini); GPT-4o og GPT-4o mini er kun tilgjengelig i US government-regioner\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/microsoft-copilot-studio/prompt-model-settings\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#11\",\n \"claim\": \"Claude Sonnet 4.5 og Claude Opus 4.5 er tilgjengelig i Copilot Studio (experimental, 200K kontekstvindu)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/microsoft-copilot-studio/prompt-model-settings\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#12\",\n \"claim\": \"Parallelle verktøykall støttes IKKE ved Minimal reasoning_effort på GPT-5; bruk Low/Medium/High for agentbruk\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/how-to/model-choice-guide?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#13\",\n \"claim\": \"Microsoft Foundry Model Router er GA og gir opptil 60% besparelse vs å alltid bruke GPT-5-familien (dokumentert av Microsoft)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#14\",\n \"claim\": \"Norway East: kun gpt-4o/gpt-4o-mini tilgjengelig som Norge-resident (Standard/Regional PTU); gpt-4.1, o-serien (o3/o4-mini/o3-mini/o1) og GPT-5-familien finnes ikke som Regional/PTU i Norway East; eneste GPT-5 med EU-residens i Norway East er gpt-5.5 via Data Zone Standard\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#15\",\n \"claim\": \"Copilot Studio prompt-builder modell-lineup: gpt-4.1-mini (default, Basic, 128K, trent t.o.m. juni 2024) | gpt-4.1 (Standard, 128K, juni 2024) | gpt-5-chat (Standard, 128K, sept 2024) | gpt-5-reasoning (Premium, 400K, sept 2024) | gpt-5.2-reasoning (Premium, 400K, okt 2024) | gpt-5.3-chat (Standard, managed) | Claude Sonnet 4.6 (Standard, 200K) | Claude Opus 4.6 (Premium, 200K) | Grok 4.1 Fast (Standard, eksperimentell/ikke anbefalt)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/microsoft-copilot-studio/prompt-model-settings\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#16\",\n \"claim\": \"GPT-5-tilgjengelighet: gpt-5 GA (begrenset, krever godkjenning aka.ms/oai/gpt5access) | gpt-5-mini GA | gpt-5-nano GA | gpt-5-chat Preview (2 versjoner) | gpt-5-codex GA (begrenset) | gpt-5-pro GA (begrenset, kun MCA-E/Default-abonnementer)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure?view=foundry-classic\"\n },\n {\n \"id\": \"ms-ai-security/cost-optimization/gpt5-gpt41-pricing-models.md#17\",\n \"claim\": \"AI Builder: prompt builder credits inkludert i premium Power Platform-planer (500 credits/bruker/mnd); default modell gpt-4.1-mini (Basic rate)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/cost-optimization/gpt5-gpt41-pricing-models.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-security/references/ai-security-engineering/entra-agent-id-zero-trust.md",
"claim_count": 13,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-security/references/ai-security-engineering/entra-agent-id-zero-trust.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-security/references/ai-security-engineering/entra-agent-id-zero-trust.md`)\n\n[\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#1\",\n \"claim\": \"Microsoft Entra Agent ID har status Public Preview (utvidet etter Ignite 2025)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/id-governance/agent-id-governance-overview\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#2\",\n \"claim\": \"Tre nye Agent ID RBAC-roller: Agent ID Administrator | Agent ID Developer | Agent Registry Administrator\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/fundamentals/whats-new-ignite-2025\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#3\",\n \"claim\": \"Copilot Studio (preview, aktiveres per miljø i Power Platform Admin Center) oppretter automatisk Entra Agent ID for hver ny agent, knyttet til Microsoft Copilot Studio agent identity blueprint med Blueprint ID 25664c89-cea5-4ab6-b924-a54fd8a19ae0\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-use-entra-agent-identities\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#4\",\n \"claim\": \"En agentidentitet er en spesialisert service principal med object ID og app ID alltid lik; ingen passord eller credentials — autentiseres kun via access tokens utstedt til plattformen agenten kjører på\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/agent-id/identity-platform/what-is-agent-id\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#5\",\n \"claim\": \"Conditional Access for agenter krever Microsoft Entra ID P1 eller P2 og Microsoft Agent 365-lisens per bruker (lisenshåndhevelse kommer snart); nettverkskontroller for agenter krever Microsoft Entra Internet Access\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/identity/conditional-access/agent-id\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#6\",\n \"claim\": \"Agent risk (high/medium/low) fra Identity Protection er tilgjengelig som CA-condition; «Agent execution environments»-condition er i Preview for managed endpoints (Windows 365 Cloud PCs for Agents)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/identity/conditional-access/agent-id\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#7\",\n \"claim\": \"CA gjelder IKKE for: blueprint sine Graph-kall (oppretting av agentidentitet/agent user); intermediary token exchange på AAD Token Exchange Endpoint: Public (Resource ID fb60f99c-7a34-4190-8149-302f77469936); når Security defaults er aktivert; eller når agent bruker API-nøkkel\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/identity/conditional-access/agent-id\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#8\",\n \"claim\": \"Blokkerte rettigheter for agentidentiteter: rollene Global Administrator, Privileged Role Administrator, User Administrator; Graph-tillatelsene Application.ReadWrite.All, RoleManagement.ReadWrite.All, User.ReadWrite.All, Directory.AccessAsUser.All\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/agent-id/identity-professional/authorization-agent-id\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#9\",\n \"claim\": \"Foundry: første agent i et prosjekt oppretter automatisk standard agent identity blueprint + standard agentidentitet; upubliserte agenter deler prosjektidentitet; publisering oppretter dedikert blueprint + unik identitet og RBAC må tildeles på nytt\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/agents/concepts/agent-identity?view=foundry\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#10\",\n \"claim\": \"Opt-out fra Entra Agent Identity per miljø i Copilot Studio er midlertidig — Microsoft vil gjøre det obligatorisk for alle nye agenter i fremtiden\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-use-entra-agent-identities\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#11\",\n \"claim\": \"Statustabell: Entra Agent ID kjerne/Agent Registry/Identity Protection for agenter/Global Secure Access for agenter = Public Preview (Frontier-program); Foundry-integrasjon Public Preview (alle Foundry-brukere); Conditional Access for agenter Public Preview; AI Prompt Shield nytt via Entra Internet Access; App Service/Azure Functions agent identity og Teams Developer Portal agent blueprints nye\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/entra/fundamentals/whats-new-ignite-2025\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#12\",\n \"claim\": \"Full Entra Agent ID-funksjonalitet krever deltakelse i Microsoft Frontier-programmet og M365 Copilot-lisens (aktiveres via M365 admin center → Copilot → Settings → User access → Copilot Frontier); Foundry-integrert agentidentitet er tilgjengelig for alle Foundry-brukere uten Frontier\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/microsoft-agent-365/admin/capabilities-entra\"\n },\n {\n \"id\": \"ms-ai-security/ai-security-engineering/entra-agent-id-zero-trust.md#13\",\n \"claim\": \"Entra Agent ID er en del av Microsoft Agent 365, Microsofts kontrollplan for agenter på tvers av virksomheten\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/microsoft-agent-365/admin/capabilities-entra\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-security/references/ai-security-engineering/entra-agent-id-zero-trust.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md",
"claim_count": 10,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#1\",\n \"claim\": \"Commitment Tier-støttede tjenester: Speech to Text (Standard) | Text to Speech (Neural) | Text Translation (Standard) | Language Understanding (Text Requests) | Azure Language (Sentiment/Key Phrase/Language Detection/NER) | Vision OCR | Document Intelligence (Custom/Invoice)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/commitment-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#2\",\n \"claim\": \"Commitment Tier forpliktelsesperiode: kalendermåned (web/connected containers) eller kalenderår (disconnected containers)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/commitment-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#3\",\n \"claim\": \"Commitment Tier: auto-renewal valgfritt og kan endres frem til midnatt UTC siste dag i måneden; ikke refunderbar etter kjøp; overage faktureres til samme sats som tier\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/commitment-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#4\",\n \"claim\": \"Commitment Tier kan IKKE brukes med multi-service Cognitive Services-ressurs; krever dedikert single-service ressurs\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/commitment-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#5\",\n \"claim\": \"PTU deployment-typer: Regional Provisioned (data i én region) | Data Zone Provisioned (data innenfor data zone, f.eks. EU/US) | Global Provisioned (global lastbalansering)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#6\",\n \"claim\": \"PTU-reservasjoner: 1-måneds eller 1-års Azure Reservations; ingen 3-årstermin for Foundry PTU; rabatt varierer per modellfamilie og term\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput-billing\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#7\",\n \"claim\": \"For GPT-4o og nyere modeller er TPM per PTU satt separat for input- og output-tokens\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#8\",\n \"claim\": \"Azure OpenAI har ingen hard limit-funksjonalitet (i motsetning til OpenAI); automatisering via Action Groups krever custom utvikling\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/concepts/manage-costs\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#9\",\n \"claim\": \"API Management AI gateway styrer også Anthropic Messages API (v2-tiers) og Google Vertex AI, tilbyr unified model API (preview), og kan integreres direkte i Microsoft Foundry (preview) for sentral token-/kostnadsgovernance\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#10\",\n \"claim\": \"PTU er kun for Azure OpenAI — ikke for øvrige Azure AI Services\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-enterprise-architecture.md",
"claim_count": 16,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-enterprise-architecture.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-enterprise-architecture.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#1\",\n \"claim\": \"Azure AI Services (Foundry Models) deployment-typer: Global Standard (automatisk routing til regioner med kapasitet) | Data Zone Standard (processing innenfor geografisk sone US/EU) | Regional Standard (én spesifikk region) | Provisioned PTU (dedikert kapasitet, SLA på latency)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#2\",\n \"claim\": \"Azure API Management backend pool: inntil 30 backends per pool\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#3\",\n \"claim\": \"APIM AI gateway medierer flere LLM-skjemaer: OpenAI Chat Completions/Responses, Anthropic Messages API (API Management v2-tiers), Google Vertex AI, samt Foundry-modeller og ikke-Microsoft-leverandører som Amazon Bedrock; styrer også MCP-servere og A2A-agent-APIer som backends\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#4\",\n \"claim\": \"APIM Unified model API (ett OpenAI-kompatibelt endepunkt med automatisk format-oversettelse over flere backends) er i preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#5\",\n \"claim\": \"AI gateway i Microsoft Foundry (governance av modeller, agenter og verktøy fra Foundry-miljøet: token-kvoter, throttling, content safety, telemetri) er i preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#6\",\n \"claim\": \"APIM semantisk caching bruker Azure Managed Redis (RediSearch-kompatibel) med llm-semantic-cache-store/llm-semantic-cache-lookup-policyer\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/genai-gateway-capabilities\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#7\",\n \"claim\": \"Azure OpenAI 429-responsens Retry-After-header kan være opptil 24 timer\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#8\",\n \"claim\": \"Azure AI Search zone redundancy krever Standard tier eller høyere + minimum 3 replicas; ingen built-in disaster recovery (krever manuell gjenoppbygging eller support-kontakt)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/reliability/reliability-ai-search\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#9\",\n \"claim\": \"Azure AI Search har ingen native multi-region replication — separat service per region, geo-replication må implementeres selv\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/reliability/reliability-ai-search\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#10\",\n \"claim\": \"Cosmos DB continuous backup gir 7-day PITR (point-in-time restore) for Foundry Agent Service disaster recovery\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/architecture/ai-ml/architecture/baseline-microsoft-foundry-chat\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#11\",\n \"claim\": \"Ingen Azure OpenAI-region i Norge per 2026-02; nærmeste regioner er Sweden Central og West Europe\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://azure.microsoft.com/global-infrastructure/services/\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#12\",\n \"claim\": \"APIM-tiers: Developer (1M calls, ingen SLA) | Basic (1M calls, SLA, 1 unit max) | Standard (10M calls, multi-region, 4 units) | Premium (unlimited, multi-region, VNet, 10+ units)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#13\",\n \"claim\": \"APIM circuit breaker er ikke tilgjengelig i Consumption tier; Standard tier minimum for AI gateway\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/api-management/backends\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#14\",\n \"claim\": \"Caching (prompt caching) for Azure OpenAI omtales som planlagt feature (ikke tilgjengelig ennå)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#15\",\n \"claim\": \"Multi-Backend Gateway-guiden bekrefter fire gateway-topologier: (1) multiple model deployments i single instance, (2) multiple instances same region/single subscription, (3) multiple instances same region/multiple subscriptions, (4) multiple instances multi-region; tagger 'Foundry Tools' og 'Azure OpenAI in Foundry Models'\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/azure-openai-gateway-multi-backend\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-enterprise-architecture.md#16\",\n \"claim\": \"Azure OpenAI er del av Foundry Models (tidligere 'OpenAI in Azure'), nå konsolidert under Foundry Tools; Azure AI Services het tidligere Cognitive Services\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-enterprise-architecture.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-governance-compliance.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-governance-compliance.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-governance-compliance.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#1\",\n \"claim\": \"Purview Data Security Posture Management (DSPM) for AI støttes for Azure AI Services via Foundry; Enterprise AI apps-dekningen omfatter Anthropic Claude (Enterprise) i tillegg til Copilot\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/purview/ai-azure-services\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#2\",\n \"claim\": \"Purview-kapabiliteter støttet for AI Services: Compliance Manager | DSPM for AI | Data Classification | Sensitivity Labels | DLP (via Entra-registrerte apps) | Audit Logging (Unified Audit Log, fanger prompts/responses) | Data Lifecycle Management | eDiscovery\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/purview/ai-azure-services\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#3\",\n \"claim\": \"Purview for Azure AI Services (Foundry) aktiveres via Microsoft Foundry Portal ('Microsoft Purview Data Security' i Control Plane) eller via Microsoft Defender for Cloud ('Data Security for Azure AI with Microsoft Purview')\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/purview/ai-azure-services\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#4\",\n \"claim\": \"Purview-aktivering for AI Services krever pay-as-you-go billing i Purview; audit logging er inkludert i Purview-lisensen\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/purview/ai-azure-services\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#5\",\n \"claim\": \"AI Services diagnostikklogg-kategorier: Audit (API-kall, autentisering) | RequestResponse (full forespørsel/respons inkl. prompts) | AllMetrics; destinasjoner: Azure Storage, Log Analytics Workspace, Event Hub\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/diagnostic-logging\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#6\",\n \"claim\": \"Telemetry logs (objektnavn som indexer, skillsets) lagres globalt i 1,5 år for Microsoft-support (unntak fra Geography-residency)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#7\",\n \"claim\": \"Diagnostic logs standard retention: 90 dager (Log Analytics default), justerbar 30-730 dager\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#8\",\n \"claim\": \"Azure Policy-definisjon 67121cc7-ff39-4ab8-b7e3-95b84dab487d = 'Azure AI services should enable data encryption with CMK'\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#9\",\n \"claim\": \"Keys kan lagres i HSM-backed Key Vault for FIPS 140-2 Level 3\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#10\",\n \"claim\": \"Foundry støtter Microsoft Entra Agent Identity for unik identifisering av AI-agenter\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#11\",\n \"claim\": \"Agent 365 gir sentralisert visning av alle AI-agenter i organisasjonen\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ai-agents/governance-security-across-organization\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-governance-compliance.md#12\",\n \"claim\": \"Purview-lisensiering: DLP Policies krever E5/F5-lisens (eller pay-as-you-go) | Sensitivity Labels krever E3/E5 | DSPM for AI ikke i standard lisens (pay-as-you-go tilgjengelig) | Audit (Unified Log) inkludert\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-governance-compliance.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-monitoring-logging.md",
"claim_count": 10,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-monitoring-logging.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/ai-services-monitoring-logging.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#1\",\n \"claim\": \"Diagnostic settings log-kategorien Trace gjelder kun for Custom Question Answering\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/diagnostic-logging\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#2\",\n \"claim\": \"Azure AI Services rapporterer med ResourceProvider == 'MICROSOFT.COGNITIVESERVICES' i AzureDiagnostics-tabellen\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/monitor-openai\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#3\",\n \"claim\": \"Log Analytics data retention: 90 dager gratis, deretter betalt\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#4\",\n \"claim\": \"Application Insights er OpenTelemetry-kompatibel APM med AI agent monitoring for Microsoft Foundry og Copilot Studio\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/app/app-insights-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#5\",\n \"claim\": \"Application Insights har en Agents View spesialisert for AI agents (token usage, cost per session)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/app/app-insights-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#6\",\n \"claim\": \"Out-of-box dashboards i Microsoft Foundry for Azure OpenAI: HTTP Requests | Tokens-Based Usage | PTU Utilization | Fine-tuning\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/monitor-openai\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#7\",\n \"claim\": \"Azure OpenAI metric-navn: TokenTransaction | GeneratedTokens | ProcessedPromptTokens | ActiveTokens (PTU) | Requests | Http429\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/monitor-openai\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#8\",\n \"claim\": \"Application Insights: 5 GB/måned gratis data ingestion per subscription\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#9\",\n \"claim\": \"Application Insights adaptive sampling: default 5 items/sec\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/app/app-insights-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/ai-services-monitoring-logging.md#10\",\n \"claim\": \"Copilot Studio-bottar kan kobles til Application Insights for conversation analytics, LUIS intent recognition performance og QnA Maker query latency\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/ai-services-monitoring-logging.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-image-analysis.md",
"claim_count": 14,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-image-analysis.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-image-analysis.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#1\",\n \"claim\": \"Image Analysis 4.0 er generelt tilgjengelig (GA) siden november 2023, bygget på Florence foundation model fra Microsoft Research\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/whats-new\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#2\",\n \"claim\": \"Image Analysis 4.0 støtter synkron OCR, dense captions (opptil 10 regioner i bildet), people detection og smart crop\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-image-analysis\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#3\",\n \"claim\": \"Visual features i Image Analysis 4.0 med regionsrestriksjoner: Caption, Dense Captions og Smart Crops kun i visse Azure-regioner; Tags, Objects, People og Read (OCR) i alle regioner\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-image-analysis\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#4\",\n \"claim\": \"Regioner med full funksjonalitet (Caption/Dense Captions/Smart Crop): East US, West US, France Central, North Europe, West Europe, Southeast Asia, East Asia, Korea Central\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/concept-describe-images-40\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#5\",\n \"claim\": \"Content moderation (adult/racy/gory) finnes i Image Analysis 3.2 (fortsatt støttet), men er fjernet i Image Analysis 4.0 — Azure AI Content Safety anbefales i stedet\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-image-analysis\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#6\",\n \"claim\": \"Multimodal embeddings (4.0) gir 1024-dimensjonal vektor\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#7\",\n \"claim\": \"Multilingual embedding-modell støtter 102 språk (2024-02-01 API); embeddings fra engelsk-modellen er ikke kompatible med multilingual-modellen\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#8\",\n \"claim\": \"Image Analysis: maks bildestørrelse 20 MB\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-image-analysis\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#9\",\n \"claim\": \"Rate limits: 10-20 requests/sekund per tier (opp til 20 requests/sekund per Standard tier)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#10\",\n \"claim\": \"Custom Vision krever trening med minimum 30 bilder per tag\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#11\",\n \"claim\": \"Objektdeteksjon: objekter mindre enn 5 % av bildestørrelsen detekteres ikke\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/concept-object-detection-40\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#12\",\n \"claim\": \"Power Automate Azure AI Vision connector støtter Image Analysis 3.2, ikke 4.0 (per feb 2026) — HTTP action kreves for 4.0-features\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#13\",\n \"claim\": \"Azure AI Vision er tilgjengelig i Norway East og Norway West\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-image-analysis.md#14\",\n \"claim\": \"Azure AI Vision Free tier (F0): 0-5000 transaksjoner/måned gratis; Standard tier heter S1\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-image-analysis.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-ocr-processing.md",
"claim_count": 13,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-ocr-processing.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-ocr-processing.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#1\",\n \"claim\": \"Azure AI Content Understanding er GA (per 2026-04) og anbefales for OCR kombinert med semantisk analyse/layout/tabeller\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#2\",\n \"claim\": \"Azure Vision v4.0 Read: synkron API, GA, del av Image Analysis 4.0 | Azure Vision v3.2 Read: asynkron, GA men legacy, ingen videre oppdateringer etter v3.2 | Document Intelligence Read: asynkron, GA, optimalisert for teksttunge dokumenter; begge editions bruker samme OCR-motor\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-ocr\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#3\",\n \"claim\": \"OCR trykt tekst støtter flere språk inkludert engelsk, fransk, tysk, italiensk, portugisisk, spansk, kinesisk, japansk, koreansk, russisk, arabisk, hindi; skriftsystemer inkl. latinske, kyrilliske, arabiske og devanagari-tegnsett\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-ocr\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#4\",\n \"claim\": \"OCR håndskrift støttes for engelsk, kinesisk (forenklet), fransk, tysk, italiensk, japansk, koreansk, portugisisk og spansk; handwritten classification kun for latinsk alfabet\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-ocr\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#5\",\n \"claim\": \"OCR input-krav: filformater JPEG/PNG/BMP/PDF/TIFF | maks filstørrelse 500 MB (4 MB gratisnivå) | dimensjoner min 50x50 til maks 10000x10000 piksler | PDF/TIFF opptil 2000 sider (kun 2 første sider på gratisnivå) | minimum teksthøyde 12 piksler for 1024x768-bilde\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/how-to/call-read-api\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#6\",\n \"claim\": \"API-alternativer: Image Analysis 4.0 Read synkron med input JPEG/PNG/BMP/GIF | Document Intelligence Read asynkron med input PDF/TIFF/JPEG/PNG/BMP/Office-filer | Azure Vision v3.2 Read asynkron legacy med input JPEG/PNG/BMP/PDF/TIFF\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/computer-vision/overview-ocr\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#7\",\n \"claim\": \"Azure AI Search Image Analysis Skill bruker v3.2 API (legacy); for v4.0-funksjonalitet kreves custom Web API skill\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/cognitive-search-skill-image-analysis\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#8\",\n \"claim\": \"AI Builder tilbyr en Text Recognition prebuilt model som bruker Azure Vision OCR under panseret (Power Automate-integrasjon)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#9\",\n \"claim\": \"SynapseML tilbyr en ReadImage-transformator for OCR i Spark-pipelines (Microsoft Fabric/Synapse)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/synapse-analytics/machine-learning/tutorial-computer-vision-use-mmlspark\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#10\",\n \"claim\": \"OCR data retention: input-bilder og ekstrahert tekst lagres midlertidig i 48 timer (operation-location URL), deretter slettet automatisk; ingen permanent lagring av kundedata\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/computer-vision/ocr-data-privacy-security\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#11\",\n \"claim\": \"OCR-data krypteres under transit med TLS 1.2 og ved hvile (Azure Storage encryption)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/computer-vision/ocr-data-privacy-security\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#12\",\n \"claim\": \"Gratisnivåer: Azure Vision v4.0 Read OCR 5000 transaksjoner/måned gratis (Standard S1) | Document Intelligence Read 500 sider/måned gratis (Standard S0)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/azure-ai-vision-ocr-processing.md#13\",\n \"claim\": \"Read Docker container finnes for on-premises deployment (compliance/air-gapped-krav)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/azure-ai-vision-ocr-processing.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/content-understanding-multimodal-analysis.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/content-understanding-multimodal-analysis.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/content-understanding-multimodal-analysis.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#1\",\n \"claim\": \"Content Understanding status: Preview (GA for kjernefunksjoner, Limited Access for face description)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/video/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#2\",\n \"claim\": \"Tekniske begrensninger video/audio: frame sampling ~1 FPS | frame resolution 512×512 px resize | kun tale transkriberes (musikk/lydeffekter/bakgrunnsstøy ignoreres)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/video/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#3\",\n \"claim\": \"Prebuilt analyzers: prebuilt-videoSearch (RAG-optimalisert, WEBVTT-transcript + keyframes) | prebuilt-videoAnalysis (generell video-metadata) | prebuilt-documentSearch (markdown for RAG)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/video/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#4\",\n \"claim\": \"audioVisual-elementer som krever returnDetails:true: transcript phrases og camera shots; markdown content, contents collection, timing, key frames og field extraction krever det ikke; key frames og camera shots gjelder kun video\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/video/elements\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#5\",\n \"claim\": \"Keyframes samples uniformt fra hver camera shot, minimum én per shot, deterministisk utvalg; cameraShotTimesMs angir startpunkt per shot (første shot starter alltid ved 0 ms); detekterer abrupte og gradvise overganger\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/video/elements\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#6\",\n \"claim\": \"Content Safety multimodal analysis er begrenset til regionene East US og West Europe\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#7\",\n \"claim\": \"Content Understanding vs Video Indexer: CU har ikke face recognition (kun face description via Limited Access) og ikke real-time analyse; Video Indexer har celebrity + custom faces, live video streaming, observed people tracking (bounding boxes); CU prises token-basert, Video Indexer page/minute-basert\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-video-indexer/insights-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#8\",\n \"claim\": \"Content Understanding er core Foundry-tjeneste tilgjengelig via Microsoft Foundry portal, Foundry SDK (Python, .NET) med ContentUnderstandingClient-klasse, og REST API POST /contentunderstanding/analyzers/{analyzerId}:analyze\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/python/api/overview/azure/ai-contentunderstanding-readme\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#9\",\n \"claim\": \"Content Understanding kan integreres som skill i Azure AI Search indexer pipeline; CU-skill gir cross-page text location metadata, cross-page table extraction og semantic chunking (vs Document Extraction/Document Layout skills); støttede formater PDF, DOCX, XLSX, PPTX\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/multimodal-search-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#10\",\n \"claim\": \"Face description er Limited Access-feature: krever Azure support request + justification, Microsoft vurderer use case før godkjenning\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/content-understanding/data-privacy\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#11\",\n \"claim\": \"Free tier er ikke tilgjengelig for Content Understanding; krever betalt Azure subscription\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/content-understanding-multimodal-analysis.md#12\",\n \"claim\": \"Webhooks for async completion er preview-feature, ikke GA (per Feb 2026)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/content-understanding-multimodal-analysis.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-custom-models.md",
"claim_count": 11,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-custom-models.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-custom-models.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#1\",\n \"claim\": \"Document Intelligence custom models status: GA (v4.0 API, api-version 2024-11-30)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-model?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#2\",\n \"claim\": \"v4.0 (GA) ga custom neural models støtte for signaturdeteksjon (min. 5 samples), overlappende felter (complete/partial overlap) og tabell-/rad-/celle-konfidensscoring\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-neural?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#3\",\n \"claim\": \"Custom neural models: opptil 50 000 siders treningsdata, trening opptil 10 timer (10 gratis timer per måned, deretter betalt trening); minimum 5 dokumenter for å komme i gang\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-neural?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#4\",\n \"claim\": \"Treningstid: custom template 1-5 minutter | custom neural 30 min-12 timer\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-model?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#5\",\n \"claim\": \"Treningskrav: min 5 dokumenter (begge typer) | maks treningssider 500 (template) / 50 000 (neural) | maks treningsstørrelse 50 MB (template) / 1 GB (neural) | filformater PDF, JPEG/JPG, PNG, BMP, TIFF, HEIF | maks 2 000 sider per dokument (F0: 2 sider) | maks filstørrelse analyse S0 500 MB / F0 4 MB | bildedimensjoner 50×50 til 10 000×10 000 px\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/how-to-guides/build-a-custom-model?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#6\",\n \"claim\": \"Treningsbudsjett: v4.0 (2024-11-30) 10 gratis timer/måned, maks 10 timer per build, faktureres for faktisk tid (min. 30 min per jobb, krever eksplisitt maxTrainingHours); v3.1/v3.0: 20 byggeoperasjoner/måned, maks 30 minutter per build, ingen betalt trening\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-model?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#7\",\n \"claim\": \"Composed models: kombiner opptil 200 custom models til én modell-ID; Document Intelligence klassifiserer dokumentet automatisk og velger best match\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/how-to-guides/compose-custom-models?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#8\",\n \"claim\": \"Custom classifier i v4.0 støtter Office-formater (DOCX, XLSX, PPTX) og dokumentsplitting (én fil → mange dokumenter); min. 5 samples per klasse\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-classifier?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#9\",\n \"claim\": \"AI Builder custom models (Document Processing) støtter kun template models, ikke neural (per januar 2026)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#10\",\n \"claim\": \"Custom neural models kan kun trenes i utvalgte regioner (inkl. West Europe, ikke Norway East); modeller kan kopieres til andre regioner (f.eks. Norway East) for produksjon\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-neural?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-custom-models.md#11\",\n \"claim\": \"Custom models lagres i 90 dager uten kostnad\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-custom-models.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-prebuilt-models.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-prebuilt-models.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-prebuilt-models.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#1\",\n \"claim\": \"Azure AI Content Understanding er fullt GA (per 2026-04) og anbefalt startpunkt for de fleste IDP-scenarier; Document Intelligence anbefales for spesifikke prebuilt-skjemaer (regnskap, ID-dokumenter, kvitteringer)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/content-understanding/choosing-right-ai-tool#azure-document-intelligence\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#2\",\n \"claim\": \"Prebuilt-modellene støtter 27 språk\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/document-intelligence/language-support/prebuilt?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#3\",\n \"claim\": \"v4.0 (GA: 2024-11-30) introduserte nye felt: ReceiptType, TaxDetails og VAT-ekstraksjon for hotellkvitteringer\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/document-intelligence/prebuilt/receipt?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#4\",\n \"claim\": \"Financial Services prebuilt-modeller: prebuilt-invoice | prebuilt-receipt | prebuilt-bankStatement | prebuilt-creditCard | prebuilt-check | prebuilt-contract | prebuilt-payStub.us\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/document-intelligence/model-overview?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#5\",\n \"claim\": \"Identity & Tax prebuilt-modeller: prebuilt-idDocument | prebuilt-healthInsuranceCard.us | prebuilt-marriageCertificate | prebuilt-tax.us.w2 | prebuilt-tax.us.1098 | prebuilt-tax.us.1099 | prebuilt-tax.us.1040 | prebuilt-tax.us (unified)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/document-intelligence/model-overview?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#6\",\n \"claim\": \"US Mortgage prebuilt-modeller: prebuilt-mortgage.us.1003 | prebuilt-mortgage.us.1004 | prebuilt-mortgage.us.1005 | prebuilt-mortgage.us.1008 | prebuilt-mortgage.us.disclosure\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/document-intelligence/model-overview?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#7\",\n \"claim\": \"Grunnleggende modeller: prebuilt-read (OCR: tekst, linjer, ord, språkdeteksjon) | prebuilt-layout (tabeller, selection marks, seksjoner, valgfrie key-value pairs) | prebuilt-document (key-value pairs, tabeller, selection marks)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/document-intelligence/model-overview?view=doc-intel-4.0.0\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#8\",\n \"claim\": \"AI Builder bruker Document Intelligence v3.1 (ikke alltid v4.0); Premium-lisens påkrevd for AI Builder; 1M AI Builder credits inkludert i visse Power Apps/Automate-lisenser\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#9\",\n \"claim\": \"Logic Apps har native connector 'Azure AI Document Intelligence' med full API v4.0-støtte og managed identity authentication\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#10\",\n \"claim\": \"Free (F0): 500 sider/måned, 2 sider per dokument, 20 calls/min | Standard (S0): 2 000 sider per dokument, 15 TPS\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#11\",\n \"claim\": \"Standard 30-dagers oppbevaring av dokumenter i Document Intelligence (kan slettes umiddelbart etter prosessering)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md#12\",\n \"claim\": \"Custom models utløper (slutter å virke) etter 1224 måneder; krever retraining-schedule\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/document-intelligence-prebuilt-models.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/language-services-custom-text-classification.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/language-services-custom-text-classification.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/language-services-custom-text-classification.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#1\",\n \"claim\": \"Custom Text Classification/Custom NER status: GA — avvikles 31. mars 2029\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/language-service/custom-text-classification/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#2\",\n \"claim\": \"Custom Text Classification og Custom NER avvikles 2029-03-31; anbefalt migrering til Microsoft Foundry-modeller (prompt-basert klassifisering og NER med GPT-4o eller GPT-4.1)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/language-service/custom-text-classification/how-to/migrate-azure-openai\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#3\",\n \"claim\": \"Custom Text Classification støtter to prosjekttyper: Single label classification (ett dokument, én kategori) og Multi label classification (ett dokument, flere kategorier)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/custom-text-classification/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#4\",\n \"claim\": \"Tjenestene er tilgjengelige via Microsoft Foundry portal (ai.azure.com) og via REST API/SDK-er for Python, C#, Java og JavaScript\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/custom-text-classification/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#5\",\n \"claim\": \"Standard NER dekker 18+ entity-typer out-of-the-box (person, location, org, quantity, datetime, etc.)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#6\",\n \"claim\": \"Language resource må ha 'Custom text classification & custom named entity recognition'-feature aktivert, som krever storage account-tilkobling (og managed identity)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/custom-named-entity-recognition/how-to/create-project\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#7\",\n \"claim\": \"Custom Text Classification kan ikke integreres direkte i M365 Copilot (Copilot bruker forhåndstrente modeller); workaround via Copilot Studio-bot eller Power Automate-flow\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#8\",\n \"claim\": \"Tekst sendt til runtime-API logges ikke, men respons caches i 15 min (kan deaktiveres med loggingOptOut: true)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#9\",\n \"claim\": \"'Text record' = inntil 1000 characters; lengre tekster teller som flere records (2500 characters = 3 records)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#10\",\n \"claim\": \"Free tier (F0): 5000 text records per måned for Prediction API | trening ubegrenset | maks 1 deployment per prosjekt\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#11\",\n \"claim\": \"Batch API: opptil 10 dokumenter per request\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-custom-text-classification.md#12\",\n \"claim\": \"Custom Classification støtter 100+ språk (multilingual-funksjonalitet)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/custom-text-classification/language-support\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/language-services-custom-text-classification.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/language-services-question-answering.md",
"claim_count": 15,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/language-services-question-answering.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/language-services-question-answering.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#1\",\n \"claim\": \"Custom Question Answering status: GA — avvikles 31. mars 2029\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/language-service/question-answering/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#2\",\n \"claim\": \"CQA avvikles 2029-03-31; anbefalt migrering til Microsoft Foundry-baserte løsninger: Agentic Retrieval, Azure AI Search + GPT-4o RAG-pipeline, eller AI Foundry Knowledge Retrieval\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/language-service/question-answering/how-to/migrate-qnamaker-to-question-answering\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#3\",\n \"claim\": \"CQA er etterfølgeren til QnA Maker, som ble retired mars 2025; migrering via export/import av TSV-filer\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/language-service/question-answering/how-to/migrate-qnamaker-to-question-answering\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#4\",\n \"claim\": \"CQA støtter 53 språk inkludert norsk; transformer-baserte rankeringsmodeller\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/language-support\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#5\",\n \"claim\": \"CQA krever to ressurser: Language resource (S-tier anbefalt for produksjon) + Azure AI Search (S1-tier for 10 TPS throughput-cap)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/concepts/azure-resources\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#6\",\n \"claim\": \"Index-allokering i Azure AI Search: N-1 indexes for single-language-prosjekter (14 prosjekter på tier med 15 indexes), N/2 for multi-language (7 prosjekter); siste index reservert for authoring/testing\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/concepts/azure-resources\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#7\",\n \"claim\": \"SDK-er: .NET Azure.AI.Language.QuestionAnswering (runtime, authoring-pakke tilgjengelig) og Python azure-ai-language-questionanswering (runtime og authoring)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/quickstart/sdk\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#8\",\n \"claim\": \"Ingen hard limit på antall dokumenter per prosjekt; maks dokumentstørrelse 25 GB per S-tier prosjekt\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#9\",\n \"claim\": \"Chit-chat personality datasets: Professional | Friendly | Witty | Caring | Enthusiastic\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#10\",\n \"claim\": \"QnA Maker native integration i Copilot Studio (Power Virtual Agents) er deprecated; må bruke REST API-tilkobling\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#11\",\n \"claim\": \"CQA er ikke en native data source for M365 Copilot (ingen direkte integrasjon)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#12\",\n \"claim\": \"Cross-region replication er ikke tillatt i CQA; Azure AI Search må være i samme region som Language resource\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/concepts/azure-resources\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#13\",\n \"claim\": \"Azure AI Search-kapasitet for CQA: Free 3 indexes / 50 MB / 2 KBs | Basic 15 indexes / 2 GB / 14 KBs (single-language) | S1 50 indexes / 25 GB / 49 KBs\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/concepts/azure-resources\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#14\",\n \"claim\": \"Language resource med CQA: Free (F0) 5000 text records/måned | Standard (S) ubegrensede hosted calls\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-question-answering.md#15\",\n \"claim\": \"CQA har hard cap på 10 TPS (transactions per second); overskridelse gir HTTP 429\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/concepts/azure-resources\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/language-services-question-answering.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/language-services-text-analytics.md",
"claim_count": 10,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/language-services-text-analytics.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/language-services-text-analytics.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#1\",\n \"claim\": \"Sentiment Analysis, Opinion Mining og Custom Text Classification avvikles 2029-03-31; PII Detection, Key Phrase Extraction og Language Detection er ikke berørt; migrasjonssti er Microsoft Foundry-modeller\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/sentiment-opinion-mining/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#2\",\n \"claim\": \"Azure AI Language text analytics-kjernekapabiliteter: Sentiment Analysis (med opinion mining) | Key Phrase Extraction | Named Entity Recognition | Language Detection; tilgjengelig som REST API, SDK (C#, Java, Python, JavaScript) og Docker-containere\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/sentiment-opinion-mining/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#3\",\n \"claim\": \"Tjenesten er stateless (ingen data lagres); batch-resultater er tilgjengelige i 24 timer før automatisk sletting\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#4\",\n \"claim\": \"Key Phrase Extraction støtter 94 språk (inkl. norsk, samisk, finsk, svensk, dansk); introduksjonen sier 94+ språk\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/key-phrase-extraction/language-support\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#5\",\n \"claim\": \"NER-entitetskategorier: Person/PersonType | Organization (Medical/Sports/StockExchange) | Location (City, CountryRegion, State, GPE, Airport, Continent) | DateTime | Quantity (Number, Percentage, Currency, Age, Temperature, Speed, Weight, Volume, Area, Length) | Event | Contact (Email, PhoneNumber, URL, IpAddress, Address) | Product/ComputingProduct | Skill, Information\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-service/named-entity-recognition/concepts/named-entity-categories\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#6\",\n \"claim\": \"Docker-containere støtter Sentiment, Language Detection, Key Phrase, Custom NER og Text Analytics for Health\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#7\",\n \"claim\": \"Docker-containere krever Standard S tier; Free F0 tier støttes ikke for containere\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#8\",\n \"claim\": \"Maks 5000 tegn per text record; større dokumenter må splittes\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#9\",\n \"claim\": \"Azure Norway-datasentre (Oslo, Stavanger) støtter ikke Language Services per 2026-02; nærmeste regioner er West Europe (Nederland) og North Europe (Irland)\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/language-services-text-analytics.md#10\",\n \"claim\": \"Free F0 tier gir 5000 gratis text records per måned\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/language-services-text-analytics.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/speech-services-speaker-recognition.md",
"claim_count": 8,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/speech-services-speaker-recognition.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/speech-services-speaker-recognition.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#1\",\n \"claim\": \"Speaker Recognition (alle features) er Limited Access: registrering kreves, kun Microsoft-managed customers, personlige e-postadresser avvises, godkjenning tar 5-10 virkedager\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/cognitive-services-limited-access\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#2\",\n \"claim\": \"Text-dependent verification bruker 10 forhåndsdefinerte engelske passphrases; similarity threshold >= 0.5 (kombinert voice + tekst); text-independent threshold >= 0.5 (kun voice)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speaker-recognition-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#3\",\n \"claim\": \"Speaker Identification: maks 50 kandidat-speakers per request\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speaker-recognition-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#4\",\n \"claim\": \"Speaker Identification-respons: 1 identified ID + 5 top-ranked IDs med scores; default threshold 0.5 (kan overstyres); returnerer \\\"0\\\"-string hvis ingen score >= 0.5\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/rest/api/speakerrecognition/\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#5\",\n \"claim\": \"Speaker Recognition-API-ene er ikke designet for liveness detection (levende person vs. opptak/imitasjon); replay attack-mitigering må implementeres separat\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speaker-recognition-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#6\",\n \"claim\": \"Speaker Recognition/Speech-ressurs tilgjengelig i Norway East (Oslo) og West Europe; voice signatures lagres i samme region som Speech resource\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#7\",\n \"claim\": \"Teams Intelligent Speaker krever Teams Rooms Pro (ikke Standard/Basic); Copilot speaker attribution krever Teams Premium eller Copilot-lisens\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoftteams/rooms/voice-recognition\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speaker-recognition.md#8\",\n \"claim\": \"Speech SDK er gratis, open source med MIT-lisens\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/speech-services-speaker-recognition.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/speech-services-speech-to-text.md",
"claim_count": 14,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/speech-services-speech-to-text.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/speech-services-speech-to-text.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#1\",\n \"claim\": \"Speech to text har tre hovedmodi: real-time transcription (live/streaming), fast transcription (synkron), batch transcription (asynkron)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#2\",\n \"claim\": \"Fast transcription: lydfiler maks 5 timer og maks 500 MB; output kun display form\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/fast-transcription-create\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#3\",\n \"claim\": \"Batch transcription: asynkron prosessering 30 min - 24 timer (best-effort scheduling); output JSON med lexical + display form; ubegrenset filstørrelse\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#4\",\n \"claim\": \"Custom speech-modeller krever deployment til custom endpoint, bortsett fra ved batch transcription; modeller utløper etter en definert periode (model lifecycle)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/custom-speech-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#5\",\n \"claim\": \"Speaker diarization: maksimalt 2 kanaler hvis diarization er aktivert; støttes ikke på tvers av flere kanaler samtidig\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#6\",\n \"claim\": \"Multi-lingual transcription er preview og støtter 15 språk: de-DE, en-AU, en-CA, en-GB, en-IN, en-US, es-ES, es-MX, fr-CA, fr-FR, it-IT, ja-JP, ko-KR, pt-BR, zh-CN\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#7\",\n \"claim\": \"Phrase list er tilgjengelig i fast transcription API version 2025-10-15\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/fast-transcription-create\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#8\",\n \"claim\": \"Støttede lydformater: WAV, MP3, OPUS/OGG, FLAC, WMA, AAC, ALAW (WAV), MULAW (WAV), AMR, WebM, SPEEX\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#9\",\n \"claim\": \"Speech SDK støtter C#, Python, Java, JavaScript, C++, Go; Speech CLI heter spx; Speech Studio på speech.microsoft.com\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#10\",\n \"claim\": \"Batch transcription API kan bruke Whisper model; Azure OpenAI tilbyr også Whisper med andre pricing og capabilities enn Azure Speech batch transcription\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#11\",\n \"claim\": \"Batch best practices: send ~1000 filer per Transcription_Create request; poll status maks 1 gang per minutt (ideelt hvert 5-10 min)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#12\",\n \"claim\": \"Custom models bygget på base models fra oktober 2023 eller senere koster penger å trene; tidligere base models er gratis å trene\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/custom-speech-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#13\",\n \"claim\": \"Azure Speech støtter West Europe og North Europe regions; audio og transkripsjondata kan lagres i EU\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-speech-to-text.md#14\",\n \"claim\": \"Logging data i Microsoft-owned storage slettes etter 30 dager; customer-owned storage kontrolleres av kunden\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/speech-services-speech-to-text.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/speech-services-text-to-speech.md",
"claim_count": 14,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/speech-services-text-to-speech.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/speech-services-text-to-speech.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#1\",\n \"claim\": \"Azure Neural TTS og Custom Neural Voice er begge GA og aktivt vedlikeholdt per 2026-04; Custom Neural Voice Pro tilbyr ytterligere enterprise-tilpasning\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/custom-neural-voice\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#2\",\n \"claim\": \"TTS tilbyr over 400 stemmer på 140+ språk og dialekter\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=tts\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#3\",\n \"claim\": \"Hvert standard neural voice-modell er tilgjengelig i 24 kHz og høy-fidelitet 48 kHz; output kan opp-/ned-samples\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#4\",\n \"claim\": \"en-US-AvaMultilingualNeural støtter 91 locales\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=tts\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#5\",\n \"claim\": \"OpenAI TTS voices er tilgjengelig via Azure Speech i regionene North Central US og Sweden Central\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#6\",\n \"claim\": \"Custom Neural Voice er en Limited Access-funksjon som krever søknad (intake form aka.ms/customneural) og godkjenning fra Microsoft\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/custom-neural-voice\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#7\",\n \"claim\": \"Batch Synthesis REST API er preview; for lange lydfiler (>10 min); real-time synthesis har <10 min audio-lengde og SDK-støtte, batch er REST API only\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-synthesis\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#8\",\n \"claim\": \"Custom Neural Voice training: 20-40 compute hours training (single-style), multi-style ca. 90 compute hours, fakturerings-cap 96 compute hours; 48kHz output etter engine upgrade\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/custom-neural-voice\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#9\",\n \"claim\": \"Billable characters: alle tegn teller inkl. mellomrom/tegnsetting; SSML markup teller unntatt <speak> og <voice> tags; kinesiske tegn (også kanji/hanja/hanzi) = 2 billable characters\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#10\",\n \"claim\": \"TTS er ikke native i M365 Copilot per januar 2026; integreres via custom connectors (Power Automate)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#11\",\n \"claim\": \"Azure OpenAI TTS API støtter tts-1 og tts-1-hd modeller med stemmene alloy, echo, fable, onyx, nova, shimmer\",\n \"claim_type\": \"version\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#12\",\n \"claim\": \"Voice Live API kombinerer STT, LLM og TTS i én WebSocket-forbindelse; støtter real-time avatar synthesis\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#13\",\n \"claim\": \"Norsk bokmål (nb-NO) har stemmene nb-NO-PernilleNeural (F) og nb-NO-FinnNeural (M); nynorsk og samisk er ikke støttet\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=tts\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/speech-services-text-to-speech.md#14\",\n \"claim\": \"Free tier (F0): 5 audio requests/month og 0.5M characters/month; S0 tier for produksjon\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://azure.microsoft.com/pricing/details/cognitive-services/speech-services/\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/speech-services-text-to-speech.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/translator-custom-neural-models.md",
"claim_count": 11,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/translator-custom-neural-models.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/translator-custom-neural-models.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#1\",\n \"claim\": \"Custom Translator (Azure Translator) har status GA\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/custom-translator/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#2\",\n \"claim\": \"Custom Translator-modell gir typisk 5-10 BLEU-poeng forbedring sammenlignet med baseline-modellen\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/custom-translator/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#3\",\n \"claim\": \"Training krever minimum 10,000 parallelle setninger (anbefalt 50,000+); tuning og testing anbefalt 2,500 setninger hver (auto-genereres hvis ikke levert)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/custom-translator/concepts/model-training\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#4\",\n \"claim\": \"Dictionary-only training er mulig uten 10,000 parallelle setninger, men gir ingen BLEU-score; phrase dictionary er case-sensitive 100% match, sentence dictionary case-insensitive\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/custom-translator/concepts/model-training\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#5\",\n \"claim\": \"Støttede filformater: TMX, XLIFF, XLF, DOCX, LCL, HTML, HTM, TXT (UTF-8/UTF-16), PDF, XLSX, ALIGN (pre-aligned); arkiver ZIP, GZ, TGZ; ZIP-navnekonvensjon {document_name}_{language_code}\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/custom-translator/concepts/document-formats-naming-convention\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#6\",\n \"claim\": \"Custom-modeller brukes via Translator Text API v3 (api-version=3.0) med category-parameter mot api.cognitive.microsofttranslator.com; CategoryID-format {WorkspaceID}-{ProjectLabel}-{CategoryCode}\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/translator/custom-translator/how-to/translate-with-custom-model\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#7\",\n \"claim\": \"Custom Translator er tilgjengelig i Microsoft Foundry (klassisk portal) for no-code workflows med document upload, BLEU-sammenligning og deployment management\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#8\",\n \"claim\": \"Document Translation (asynkron batch) kan bruke custom-modeller med samme CategoryID-parameter; krever Azure Blob Storage for input/output; Speech Service kan også bruke Custom Translator CategoryID for speech translation\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#9\",\n \"claim\": \"Microsoft Translator har EU Data Boundary-sertifisering; baseline NMT-modeller er trent på global data, custom-modeller bruker kun kundens data\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#10\",\n \"claim\": \"Custom Translator støttes ikke på Free (F0) tier (kun baseline-modeller); Standard (S1) gir full støtte med ubegrenset antall workspaces, projects og modeller\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-custom-neural-models.md#11\",\n \"claim\": \"Custom Translator støtter 60+ språkpar; språkparet må inkludere engelsk som source eller target\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/language-support\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/translator-custom-neural-models.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/azure-ai-services/translator-document-translation.md",
"claim_count": 16,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/azure-ai-services/translator-document-translation.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/azure-ai-services/translator-document-translation.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#1\",\n \"claim\": \"Azure Translator støtter tekst- og dokumentoversettelse på tvers av over 135 språk og dialekter\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#2\",\n \"claim\": \"Batch document translation støtter bildefilformater (.jpeg, .png, .bmp, .webp) — nyhet desember 2025, preview (API-versjon 2025-12-01-preview)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/reference/start-batch-translation#translate-image-files\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#3\",\n \"claim\": \"Støttede batch-dokumentformater: .docx|.xlsx|.pptx|.msg|.pdf (OCR for scannede)|.html|.htm|.csv|.tsv|.mhtml|.md|.xlf|.odt|.ods|.odp\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#4\",\n \"claim\": \"Legacy-formater .doc, .xls, .ppt konverteres automatisk til .docx, .xlsx, .pptx ved oversettelse\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#5\",\n \"claim\": \"Funksjonsmatrise batch vs sync: multi-file translation, large file support, image text translation og multi-language docs kun batch; immediate response kun sync; custom glossaries, custom models og auto language detect i begge\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#6\",\n \"claim\": \"Oversettelse av bildetekst i .docx/.pptx (batch) krever Azure AI Services multi-service resource (ikke standalone Translator) + parameter \\\"translateTextWithinImage\\\": true i options-feltet; tilleggskostnad basert på Azure Vision-prising\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/reference/start-batch-translation#translate-image-files\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#7\",\n \"claim\": \"Batch-response inkluderer feltene totalImageScansSucceeded og totalImageScansFailed for monitorering\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/reference/start-batch-translation#translate-image-files\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#8\",\n \"claim\": \"Synchronous single-file translation er egnet for filer < 40 MB (batch anbefalt for filstørrelse > 40 MB)\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/service-limits#document-translation\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#9\",\n \"claim\": \"Document Translation tilgjengelig via Foundry (classic) portal som no-code playground; Foundry (new) støtter kun forhåndsdefinerte språk med sample-dokumenter (ikke egne filer)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#10\",\n \"claim\": \"Power Automate har Translator v3 connector som støtter både text og document translation\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#11\",\n \"claim\": \"Data residency: oversettelse prosesseres i nærmeste Azure-region (Global) eller spesifisert geografi (Americas, Asia Pacific, Europe); Europa-eksempler France Central og West Europe\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/overview\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#12\",\n \"claim\": \"Microsoft lagrer ikke brukerdata etter oversettelse (dokumentert i Transparency Note)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/foundry/responsible-ai/translator/transparency-note\"\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#13\",\n \"claim\": \"Norway-region finnes ikke for Translator — bruk West Europe eller North Europe for geografisk nærhet\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#14\",\n \"claim\": \"Translator Container (disconnected deployment) finnes for air-gapped miljøer\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#15\",\n \"claim\": \"Volume Discount Plans C2, C3, C4, D3 tilgjengelig for store volumer (commitment-basert)\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/azure-ai-services/translator-document-translation.md#16\",\n \"claim\": \"totalCharacterCharged returneres i response headers for cost tracking\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/ai-services/translator/document-translation/how-to-guides/use-rest-api-programmatically\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/azure-ai-services/translator-document-translation.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/data-engineering/data-factory-ai-pipelines.md",
"claim_count": 16,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/data-engineering/data-factory-ai-pipelines.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/data-engineering/data-factory-ai-pipelines.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#1\",\n \"claim\": \"Data Factory ekstraherer data fra 170+ kilder\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/fabric/data-factory/data-factory-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#2\",\n \"claim\": \"Copilot-drevet pipeline-design (natural language → pipeline-logikk) er kun tilgjengelig i Fabric Data Factory, ikke Azure Data Factory\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/fabric/data-factory/data-factory-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#3\",\n \"claim\": \"Status: GA for både Azure Data Factory og Fabric Data Factory\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/fabric/data-factory/data-factory-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#4\",\n \"claim\": \"Fabric Data Factory har native Azure Machine Learning Activity med batch endpoint + pipeline (v1) support\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#5\",\n \"claim\": \"Azure ML Update Resource activity støtter kun batch endpoints; online endpoints krever ARM templates eller Azure ML SDK via Notebook Activity\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/data-factory/transform-data-machine-learning-service\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#6\",\n \"claim\": \"Fabric Data Factory har kun AzureMLExecutePipeline (ikke AzureMLUpdateResource — update resource må gjøres via REST API)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/fabric/data-factory/migrate-planning-azure-data-factory\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#7\",\n \"claim\": \"Fabric tilbyr Apache Airflow Job (code-first Python DAGs) som alternativ til Pipeline, med Airflow providers apache-airflow-providers-microsoft-azure\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#8\",\n \"claim\": \"Copilot Studio har ikke native Data Factory connector (per Feb 2026)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#9\",\n \"claim\": \"Data Factory kan ikke direkte kalle AI Builder-modeller (per Feb 2026)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#10\",\n \"claim\": \"Azure Data Factory støtter Norway East/West regions (metadata lagres i region)\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#11\",\n \"claim\": \"Fabric capacity er tilgjengelig i Norway West/East per Q1 2026 (OneLake-data residency følger capacity-region)\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#12\",\n \"claim\": \"Fabric Data Factory CU-forbruk: Data movement (Copy) 1.5 CU/hour per intelligent throughput unit | Orchestration 0.0056 CU per activity run\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/fabric/data-factory/pricing-pipelines\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#13\",\n \"claim\": \"Fabric Capacity SKUs: F2 = 2 CU/sekund | F64 = 64 CU/sekund | F256 = 256 CU/sekund\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://azure.microsoft.com/en-us/pricing/details/microsoft-fabric/\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#14\",\n \"claim\": \"Azure Data Factory er ikke HIPAA BAA-compliant (per Feb 2026); Azure Synapse Pipelines er HIPAA-certified (samme teknologi)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#15\",\n \"claim\": \"Microsoft tilbyr PowerShell-modulen Microsoft.FabricPipelineUpgrade for migrering av pipelines fra Azure Data Factory til Fabric Data Factory\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/fabric/data-factory/migrate-planning-azure-data-factory\"\n },\n {\n \"id\": \"ms-ai-engineering/data-engineering/data-factory-ai-pipelines.md#16\",\n \"claim\": \"Azure ML SDK v1 er deprecated per 2025-03-31 (data ingestion with ADF-dokumentasjonen gjelder SDK v1)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/how-to-data-ingest-adf\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/data-engineering/data-factory-ai-pipelines.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md",
"claim_count": 18,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#1\",\n \"claim\": \"Built-in LLM judges: Groundedness, Relevance, Coherence scorer på 1-5 skala; safety evaluators Violence, Sexual, Self-harm, Hate/Unfairness på 0-7 skala\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/develop/evaluate-sdk\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#2\",\n \"claim\": \"Agentic evaluators i Azure AI Evaluation: IntentResolutionEvaluator, TaskAdherenceEvaluator, ToolCallAccuracyEvaluator\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/develop/evaluate-sdk\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#3\",\n \"claim\": \"Max 20 scorers per experiment i MLflow production monitoring\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/mlflow3/genai/eval-monitor/\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#4\",\n \"claim\": \"MLflow 3 evaluation krever SDK mlflow[databricks]>=3.1\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/mlflow3/genai/eval-monitor/\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#5\",\n \"claim\": \"MLflow 3 built-in judges + ground truth-krav: RelevanceToQuery nei | RetrievalRelevance nei | RetrievalGroundedness nei | Safety nei | Correctness ja | RetrievalSufficiency ja | ToolCallCorrectness ja | ToolCallEfficiency nei | Guidelines nei | ExpectationsGuidelines nei (guidelines i expectations)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/mlflow3/genai/eval-monitor/concepts/scorers\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#6\",\n \"claim\": \"MLflow 3 multi-turn (conversation-level) judges: ConversationCompleteness, UserFrustration, KnowledgeRetention, ConversationalSafety, ConversationalGuidelines, ConversationalRoleAdherence, ConversationalToolCallEfficiency\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/mlflow3/genai/eval-monitor/concepts/scorers\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#7\",\n \"claim\": \"MLflow production monitoring bruker Databricks-hosted LLM judges; EU-workspaces bruker EU-hosted modeller; ingen prompts lagres hos Azure OpenAI (Abuse Monitoring opt-out)\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/mlflow3/genai/eval-monitor/\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#8\",\n \"claim\": \"MLflow 3 erstattet Agent Evaluation SDK — migrer med mlflow.genai.*-funksjoner\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/mlflow3/genai/eval-monitor/\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#9\",\n \"claim\": \"Power Platform har ingen built-in continuous evaluation for LLM i production (gap per feb 2026)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#10\",\n \"claim\": \"Copilot Studio: ingen out-of-the-box production evaluation; Microsoft roadmap (Q2 2026) inkluderer native integrasjon med Microsoft Foundry evaluation\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#11\",\n \"claim\": \"Databricks krever Premium eller Enterprise workspace tier for Unity Catalog governance\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#12\",\n \"claim\": \"azure.ai.evaluation tilbyr AIRedTeamingAgent for automatiserte adversarial scans (attack strategies: jailbreak, prompt_injection, bias_elicitation)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/python/api/overview/azure/ai-evaluation-readme\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#13\",\n \"claim\": \"azure.ai.evaluation tilbyr PIIMaskingPreprocessor med mask_types (email, phone, ssn, name)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/python/api/overview/azure/ai-evaluation-readme\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#14\",\n \"claim\": \"azure.ai.evaluation inkluderer ViolenceEvaluator, ProtectedMaterialEvaluator og CodeVulnerabilityEvaluator\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/python/api/overview/azure/ai-evaluation-readme\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#15\",\n \"claim\": \"Continuous evaluation for AI agents i Microsoft Foundry er preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/how-to/continuous-evaluation-agents\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#16\",\n \"claim\": \"Prompt Flow pensjoneres 2027-04-20 — migrer til Microsoft Agent Framework; Azure ML model monitoring for generative AI applications er preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/machine-learning/prompt-flow/how-to-monitor-generative-ai-applications\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#17\",\n \"claim\": \"Agent Monitoring Dashboard i Microsoft Foundry er preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry/observability/how-to/how-to-monitor-agents-dashboard\"\n },\n {\n \"id\": \"ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#18\",\n \"claim\": \"Azure AI Evaluation SDK versjon v1.17.0 (per research-dato 2026-06-19)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/python/api/overview/azure/ai-evaluation-readme\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/rag-architecture/azure-ai-search-setup.md",
"claim_count": 15,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/rag-architecture/azure-ai-search-setup.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/rag-architecture/azure-ai-search-setup.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#1\",\n \"claim\": \"Tier-grenser: Free 10 000 docs / 3 indekser | Basic 15M docs / 15 indekser | S1 60M per partition / 50 indekser | S2 120M per partition / 200 indekser | S3 240M per partition / 200 indekser | L1/L2 120M/240M per partition / 10 indekser\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#2\",\n \"claim\": \"SLA 99.9% for 2+ replicas (query) og 3+ replicas (indexing + query); Basic kan møte SLA-krav med opptil 3 replicas\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#3\",\n \"claim\": \"Semantic Ranker er tilgjengelig på alle tiers (også Free) — gratis månedlig kvote (default) + standard pay-as-you-go-plan ved bruk utover kvoten\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/semantic-search-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#4\",\n \"claim\": \"Private Link (inbound) ikke tilgjengelig på Free tier; indexer-utgående tilkoblinger med skillsets støttes ikke på Free, Basic, S1 eller S3 HD\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#5\",\n \"claim\": \"Search services opprettet etter april 2024 (andre bølge mai 2024) i utvalgte regioner har større partisjoner og høyere vector-kvoter på samme billing-rate\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#6\",\n \"claim\": \"In-place tier-bytte støttes mellom Basic og Standard (S1/S2/S3); bytte til/fra Free, L1/L2 eller S3 HD krever ny service + migrering; S3 HD er hosting mode for mange små indekser med samme per-enhet-kostnad som S3\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#7\",\n \"claim\": \"Serverless (Developer) er ny consumption-basert prismodell (CU/h + GB/mnd) i Public Preview, kun i utvalgte regioner, støtter ikke migrering til/fra andre tiers\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#8\",\n \"claim\": \"Indexer (pull-modell) støttede kilder: Azure Blob Storage, Cosmos DB, SQL Database, SharePoint Online\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-indexer-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#9\",\n \"claim\": \"Free tier: ingen SLA, maks 10K docs, delte ressurser, kun én Free-service per subscription\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-sku-tier\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#10\",\n \"claim\": \"Prompt Flow (Microsoft Foundry + Azure ML) pensjoneres 2027-04-20 — migrer til Microsoft Agent Framework (MAF)\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/azure/machine-learning/prompt-flow/migrate-prompt-flow-to-agent-framework?view=azureml-api-2\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#11\",\n \"claim\": \"AI Foundry (tidligere AI Studio) har innebygget Vector Index-node i Prompt Flow med index_type Azure AI Search\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#12\",\n \"claim\": \"Azure AI Search støtter Norway East og Norway West regioner (data lagres kun i Norge, ingen geo-replikering utenfor EU/EEA)\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#13\",\n \"claim\": \"AI Search throttler Push API ved >1000 docs per batch\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#14\",\n \"claim\": \"Hybrid search kjører full-text og vector search parallelt og merger resultater med Reciprocal Rank Fusion (RRF); støtter filtrering, faceting, sortering, scoring profiles og semantic ranking i én request\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/hybrid-search-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/azure-ai-search-setup.md#15\",\n \"claim\": \"maxTextRecallSize (preview) kontrollerer antall BM25-resultater inn til RRF-ranker: default 1000, max 10000\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/hybrid-search-overview\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/rag-architecture/azure-ai-search-setup.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/rag-architecture/rag-document-preprocessing.md",
"claim_count": 14,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/rag-architecture/rag-document-preprocessing.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/rag-architecture/rag-document-preprocessing.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#1\",\n \"claim\": \"Indexers kan kjøres on-demand eller på schedule ned til hvert 5. minutt\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-indexer-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#2\",\n \"claim\": \"Document Intelligence Layout-modellen støtter 309 trykte språk + 12 håndskrevne språk\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept/retrieval-augmented-generation\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#3\",\n \"claim\": \"Layout-modellen produserer Markdown-output, konverterer tabeller til Markdown, bevarer dokumentstruktur (headings, sections, paragraphs) og håndterer PDF-er, bilder, Office-filer og HTML i ett API-kall\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept/retrieval-augmented-generation\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#4\",\n \"claim\": \"Innebygde skills i Azure AI Search: OCR, Document Extraction, Text Merge, Text Split (størrelsesbasert), Azure Content Understanding Skill (semantisk layout-bevisst chunking, alternativ til Text Split for integrated vectorization), Document Intelligence Layout Skill, Language Detection\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/cognitive-search-concept-intro\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#5\",\n \"claim\": \"Støttede datakilder for indexers: Azure Blob Storage, Azure Data Lake Storage Gen2, Azure Cosmos DB, Azure SQL Database, SharePoint Online (preview), OneLake (Microsoft Fabric)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-indexer-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#6\",\n \"claim\": \"Azure OpenAI Embedding Skill med deploymentId text-embedding-3-large og dimensions 3072\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-how-to-integrated-vectorization\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#7\",\n \"claim\": \"Indexer batchSize har default 1\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/search-indexer-overview\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#8\",\n \"claim\": \"Document Intelligence v4.0 (2024-11-30) er GA\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept/retrieval-augmented-generation\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#9\",\n \"claim\": \"Azure AI Search lagringsgrenser: Free 50 MB / 3 indexes | Basic 2 GB / 15 indexes | Standard S1 25 GB / 50 indexes\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://azure.microsoft.com/en-us/pricing/details/search/\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#10\",\n \"claim\": \"Document Intelligence tiers: Free 500 sider/måned | Standard S0 pay-as-you-go\",\n \"claim_type\": \"sku\",\n \"evidence_url\": \"https://azure.microsoft.com/en-us/pricing/details/ai-document-intelligence/\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#11\",\n \"claim\": \"Azure OpenAI krever søknad om tilgang (compliance-vurdering)\",\n \"claim_type\": \"status\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#12\",\n \"claim\": \"imageAction: generateNormalizedImages er påkrevd for bildebehandling; maks 1000 bilder ekstraheres per dokument; kostnader påløper ved imageAction != none\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/cognitive-search-concept-image-scenarios\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#13\",\n \"claim\": \"Azure AI Search støtter tre tilnærminger til bildeinnhold i RAG: vektorisering (Azure Vision genererer søkbare vektorer), verbalisering (GenAI Prompt skill → LLM-tekstbeskrivelse), analyse/OCR (Image Analysis skill + OCR skill)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/search/cognitive-search-concept-image-scenarios\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-document-preprocessing.md#14\",\n \"claim\": \"Azure OpenAI i Norge: data residency i Norge (West Europe region for Azure AI Services)\",\n \"claim_type\": \"region\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/rag-architecture/rag-document-preprocessing.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-engineering/references/rag-architecture/rag-hallucination-mitigation.md",
"claim_count": 9,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-engineering/references/rag-architecture/rag-hallucination-mitigation.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-engineering/references/rag-architecture/rag-hallucination-mitigation.md`)\n\n[\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#1\",\n \"claim\": \"Groundedness-deteksjon: GA | groundedness correction/mitigating-feature: Preview, api-version 2024-09-15-preview\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#2\",\n \"claim\": \"Groundedness Detection-moduser: Non-Reasoning (rask, score uten forklaring) | Reasoning (detaljert forklaring av ungrounded segments) | Correction (auto-korrigering av ungrounded tekst)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#3\",\n \"claim\": \"Groundedness Detection API-parametre: domain=MEDICAL|GENERIC, task=QnA|Summarization, groundingSources (array), reasoning true/false, correction true\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#4\",\n \"claim\": \"Correction (auto-korrigering) krever Azure OpenAI GPT-4o som llmResource\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#5\",\n \"claim\": \"Azure OpenAI On Your Data grounding-parametre: strictness 1-5 | inScope true/false | top_n_documents\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#6\",\n \"claim\": \"Nødvendige ressurser: Azure OpenAI E0-tier | Azure AI Content Safety Standard tier (Groundedness Detection inkludert) | Azure AI Search S1 eller høyere for RAG-indexing\",\n \"claim_type\": \"sku\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#7\",\n \"claim\": \"Breaking change (oppdatert 2026-04): API-feltnavn endret correction → mitigating (deteksjons-modus) og correctedText → correctionText (output-felt)\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#8\",\n \"claim\": \"Groundedness Detection: kun engelsk tekst støttes; tilgjengelig i utvalgte Azure-regioner\",\n \"claim_type\": \"region\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/groundedness\"\n },\n {\n \"id\": \"ms-ai-engineering/rag-architecture/rag-hallucination-mitigation.md#9\",\n \"claim\": \"Foundry GroundednessEvaluator (azure.ai.evaluation Python SDK) returnerer groundedness-score 1-5 + groundedness_reason\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-engineering/references/rag-architecture/rag-hallucination-mitigation.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-governance/references/monitoring-observability/alerting-strategies-escalation.md",
"claim_count": 12,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-governance/references/monitoring-observability/alerting-strategies-escalation.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-governance/references/monitoring-observability/alerting-strategies-escalation.md`)\n\n[\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#1\",\n \"claim\": \"Alert processing rules: filtre støtter opptil 5 verdier per filter (logisk OR); Suppression har høyere prioritet enn Apply action groups\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/alerts-processing-rules\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#2\",\n \"claim\": \"Email-notifications: opptil 1000 mottakere per action group; rate limit 1000 emails per hour per action group\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups#service-limits-for-notifications\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#3\",\n \"claim\": \"SMS: 1 SMS per 5 min per telefonnummer | Voice call: 1 call per 5 min per telefonnummer\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups#service-limits-for-notifications\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#4\",\n \"claim\": \"Webhook retry: 5 retries med intervaller 5s/20s/5s/40s/5s + 15-min cooldown\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups#webhook\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#5\",\n \"claim\": \"Event Hub-notification: supports Private Link og NSP; global resilience regional\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#6\",\n \"claim\": \"Opptil 5 action groups per alert rule; action groups kjøres concurrent uten bestemt rekkefølge; samme action group kan brukes av flere alert rules\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#7\",\n \"claim\": \"Managed identity-støtte per action type: Automation Runbook (Automation Contributor, Role ID f353d9bd) ja | Logic App (Logic App Contributor, 87a39d53) ja | Event Hub (Event Hubs Data Sender, 2b629674) ja | Azure Function nei | Webhook nei | Secure Webhook nei | ITSM Connector nei\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#8\",\n \"claim\": \"Event Hub action: cross-tenant support opptil API-versjon 2023-09-01-preview\",\n \"claim_type\": \"version\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#9\",\n \"claim\": \"Managed Identity-støtte for Action Groups er i preview; Azure Portal legger automatisk til role assignments, PowerShell/CLI/SDK krever manuell rolletildeling\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/action-groups\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#10\",\n \"claim\": \"Dynamic thresholds i metric alerts: ML-baserte terskler som tilpasser seg automatisk\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/best-practices-alerts\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#11\",\n \"claim\": \"Managed identity for log search alert rules er påkrevd hvis query aksesserer Azure Data Explorer (ADX) eller Azure Resource Graph (ARG)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/best-practices-alerts\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/alerting-strategies-escalation.md#12\",\n \"claim\": \"Activity Log alerts, Service Health alerts og Resource Health alerts er gratis (ingen kostnad per rule/evaluering); first 10 metric alert rules free per subscription\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/azure-monitor/alerts/best-practices-alerts\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-governance/references/monitoring-observability/alerting-strategies-escalation.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-governance/references/monitoring-observability/security-and-audit-logging-ai.md",
"claim_count": 10,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-governance/references/monitoring-observability/security-and-audit-logging-ai.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-governance/references/monitoring-observability/security-and-audit-logging-ai.md`)\n\n[\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#1\",\n \"claim\": \"Resource logs er deaktivert som standard for Azure AI-tjenester, med unntak av Microsoft Foundry som har automatisk (auto-enabled) logging til Azure Monitor/Log Analytics\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/diagnostic-logging\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#2\",\n \"claim\": \"Azure OpenAI loggkategorier: Audit, RequestResponse, Trace (krever diagnostic setting, ikke enabled som standard)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/monitor-openai\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#3\",\n \"claim\": \"Azure AI Services loggkategorier: Audit, RequestResponse, AllMetrics (krever diagnostic setting)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/ai-services/diagnostic-logging\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#4\",\n \"claim\": \"Microsoft Defender for Cloud AI threat protection + AI-SPM: støtter Azure OpenAI; konfigureres separat for Foundry-ressurser; «Foundry resource» (kind=AIServices) er nytt navn på tidligere «Foundry Tools»\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/security/benchmark/azure/mcsb-v2-artificial-intelligence-security\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#5\",\n \"claim\": \"Log Analytics retention limit: maks 730 dager uten export\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#6\",\n \"claim\": \"Log Analytics retention er gratis de første 31 dagene\",\n \"claim_type\": \"tpm\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#7\",\n \"claim\": \"Azure Policy built-in policy finnes: «Diagnostic logs in Azure AI services resources should be enabled»\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/governance/policy/samples/azure-security-benchmark\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#8\",\n \"claim\": \"Azure security baseline for Azure OpenAI er basert på MCSB v1.0; produktet refereres som «Foundry Tools» i baseline-dokumentet\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/security/benchmark/azure/baselines/azure-openai-security-baseline\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#9\",\n \"claim\": \"Microsoft Foundry security baseline (MCSB v1.0): tjenesten omdøpt til «Microsoft Foundry»; Customer Lockbox ikke støttet; lokal autentisering til data plane ikke støttet; DLP/sensitive data discovery ikke støttet nativt\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/security/benchmark/azure/baselines/azure-ai-foundry-security-baseline\"\n },\n {\n \"id\": \"ms-ai-governance/monitoring-observability/security-and-audit-logging-ai.md#10\",\n \"claim\": \"Databricks Unity Catalog: sentral governance for både data og AI assets (modeller, features, lineage til kolonnenivå); tre governance-modeller (sentralisert, distribuert/federated, hybrid); audit-logging på workspace- og account-nivå + verbose audit logs; AI-genererte kommentarer støttes\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/databricks/lakehouse-architecture/data-governance/best-practices\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-governance/references/monitoring-observability/security-and-audit-logging-ai.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-governance/references/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md",
"claim_count": 9,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-governance/references/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-governance/references/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md`)\n\n[\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#1\",\n \"claim\": \"Copilot Studio flows finnes i to format: agent flows (eksisterende, Power Automate-lignende) og Workflows (nytt format med revidert visuell designer, public preview); begge kan utløses manuelt, av hendelser/agenter eller tidsplan\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio#accessibility\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#2\",\n \"claim\": \"Copilot Studio bruker Azure OpenAI GPT-modell for AI-kapabiliteter\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio#accessibility\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#3\",\n \"claim\": \"Retirement: etter utgangen av juni 2026 kan ikke Copilot Studio for Teams-appen brukes til å lage klassiske chatbots — makers omdirigeres til Copilot Studio-webappen\",\n \"claim_type\": \"status\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio#accessibility\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#4\",\n \"claim\": \"Copilot Studio innebygd tilgjengelighet: authoring canvas bygget etter Microsoft accessibility guidelines; standard navigasjonsmønstre (tastatur, skjermleser); ARIA-semantikk for adaptive card rendering; flerspråklig støtte inkludert norsk\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio#accessibility\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#5\",\n \"claim\": \"Azure AI Speech norsk språkstøtte: bokmål (nb-NO), nynorsk (nn-NO), Custom Speech for dialektvarianter; funksjoner: real-time transcription, speaker diarization, profanity filter/content moderation, batch-transkripsjon\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#6\",\n \"claim\": \"Azure AI Vision-kapabiliteter: Image Analysis API (beskrivende bildetekster), OCR, Face API (attributtgjenkjenning); innebygd content moderation og PII detection\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#7\",\n \"claim\": \"Produktnavn: Azure AI Document Intelligence (tidligere Form Recognizer)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#8\",\n \"claim\": \"Azure OpenAI Service Content Safety: automatisk filtrering av hatefullt språk/vold/seksuelt innhold, jailbreak detection, groundedness detection\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md#9\",\n \"claim\": \"Power Apps støtter skjermlesere (Narrator, JAWS, NVDA), tastaturnavigasjon uten mus og høykontrast-modus\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": null\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-governance/references/norwegian-public-sector-governance/accessibility-requirements-wcag-norway.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
},
{
"file": "skills/ms-ai-governance/references/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md",
"claim_count": 6,
"prompt": "# Per-claim groundedness judge — bake-off **v3.1** (recall-hardened over v3's 3 confirmed FNs)\n\nv3.1 of `judge-claim-prompt-v3.md`. Same blind, per-claim, one-subagent-per-file\ndesign, same three verdicts, same output schema, same evidence discipline. v3.1\nchanges only **three reasoning rules** (R1, R7, and a new R8), each fixing one of the\n**3 false negatives v3 still carried** (the judge-vs-gold disagreements G5 confirmed\nwere genuine judge misses, not stale gold).\n\n**Why v3.1 exists — and what it is NOT (transparent, not p-hacking).** The G5b\nfreshness spot-check (2026-06-30) blind-re-adjudicated v3's 4 apparent *false\npositives* against live Microsoft Learn. **All 4 turned out to be stale gold — v3\nflagged every one correctly** (`adr-template#1` \"zero permission management\" is\ncontradicted; `multi-region#2` lists retired `gpt-35-turbo`; `network-resilience#4`\noverstates \"recommended\" as \"obligatorisk\"; `vector-storage#7` cites the wrong GA\ndate). So **v3 has zero real false positives** (P = 100% on corrected gold), and the\nprecision-side \"FP-vakt\" originally planned for v3.1 is **dropped — there is nothing to\ndefend.** v3.1 is therefore a **pure recall hardening**: it tightens three rules so the\njudge catches 3 documented failure modes it currently misses, without touching the\nprecision-side rules (R2, R5, R6 and v3's R7 capability-following are unchanged).\n\n**Adoption is gated on measurement, not assertion — and the bar is now v3 (P 100% /\nR 92.9% on corrected gold), not v2.** v3 sits at the precision ceiling, so v3.1 can only\nbe adopted if it **holds P = 100% AND lifts R above 92.9%** (catches FNs without\nintroducing a single new false positive). Any new FP drops P below 100% and fails the\ngate — keep v3. Recall rules are double-edged over the full population (v3's own\nbake-off taught this), so the 45-way fan-out, not this prose, decides. v3 results stay\nfrozen; v3.1 writes to `judge-bakeoff-results-v3.1.json` and is graded against the\ncorrected `gold-correctness-set.json`.\n\n---\n\nYou are a correctness judge for Microsoft AI reference documentation. You verify\nfactual claims against **live, official Microsoft Learn** (`learn.microsoft.com`).\nBe strict and adversarial — do not give the benefit of the doubt, do not pad, do not\ninfer a value the source does not state.\n\nYou are judging claims extracted from `skills/ms-ai-governance/references/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md`. For EACH claim in the batch below,\ndecide whether the cited Microsoft Learn source **grounds** the claim.\n\n## The three verdicts (exhaustive, mutually exclusive)\n\n- **`grounded`** — you fetched a `learn.microsoft.com` page that states the claimed\n value(s). The page supports the claim. (Maps to gold `correct`.)\n- **`not_grounded`** — you fetched a `learn.microsoft.com` page that states a\n **different / contradicting / superseded** value for what the claim asserts. The\n claim disagrees with the source. (Maps to gold `outdated` + `wrong`.)\n- **`source_silent`** — you fetched the cited page (and searched as a fallback) but\n **no** `learn.microsoft.com` page states the claimed value at all. You cannot\n confirm or refute it. (Maps to gold `unsourced`.) Pricing on JS-rendered Azure\n pages typically lands here — that is expected, not a failure. **Exception: existence\n claims — see Rule R2.**\n\n## ⚠️ EXACT-VALUE RULE (inherited from v2 — still in force)\n\nA claim is `grounded` ONLY if the fetched page states the **exact** asserted value(s).\nVerifying that the page \"is about\" the SKU/model/feature is **not** enough — the\nspecific number, name, date, tier, dimension, or status must match. If the claim\nasserts value **X** and the page states a **different** value **Y** (even if adjacent\nor plausible), the verdict is **`not_grounded`**. This rule does NOT lower the bar for\n`not_grounded`: you still need a fetched quote stating the **differing** value.\n\nApplies with special force to `sku`, `taxonomy`, `version`, `tpm`, `region`, `status`.\n\n---\n\n## CALIBRATION RULES — read all eight before judging\n\nThe exact-value rule is a blunt instrument. The 8 rules below sharpen it on both\nedges: **R1R4 and R8 catch real errors** (more `not_grounded`); **R5R7 stop\nover-flagging where the core is grounded** (more correctly `grounded`). When a rule\nbelow conflicts with a literal reading of the exact-value rule, the rule below governs\n— it is the more precise standard.\n\n### Recall side — flag these as `not_grounded`\n\n**R1 — Bound understatement / overstatement (fixes FN2; v3.1 splits lower vs upper).**\nA claim may assert a **bound**. Direction matters — judge it by which side the bound\nconstrains:\n\n- **Lower bound** (\"100+\", \"200k+\", \"at least N\", \"minimum N\"): do not auto-`grounded`\n it just because the true value satisfies the inequality. Apply the **lower-bound\n policy:** if the true current value **grossly exceeds** the stated bound — roughly\n **>2× and decision-changing** — the bound materially misleads → `not_grounded`. A\n *tight* lower bound (true value within the same order of magnitude) stays `grounded`.\n *Example (FN2): \"200k+ context\" while the page states 1,047,576 (~1M) — ~5× → `not_grounded`.*\n- **Upper bound** (\"up to N\", \"opptil N\", \"maximum N\", \"no more than N\", \"as many as N\"):\n this is a **ceiling**, not a floor. The lower-bound leniency does **NOT** apply. The\n exact-value rule governs: if the live page states a current maximum **higher** than N,\n the stated ceiling is **superseded** → `not_grounded` — **regardless of ratio** (even\n a 1.1× exceedance breaks a ceiling). The claim tells the reader the limit is N when it\n is really higher. *Example (v3-FN): claim \"up to 18 underlying models\" while the page\n states 28 → the ceiling has moved → `not_grounded`.* (Only `grounded` if the true\n maximum is N or the claim's ceiling still binds.)\n\n**R2 — `source_silent` does NOT excuse an existence claim (fixes FN3, FN5).** When the\nclaim asserts that a named entity **exists / is offered / is in a list** (\"X is a\nbuilt-in judge\", \"feature Y is available\", \"tier Z exists\"), and you fetch the\nauthoritative page that *would* enumerate it and the entity is **absent**, that absence\nis **evidence the claim is wrong** — return `not_grounded`, not `source_silent`. Reserve\n`source_silent` for values a page would not be expected to enumerate (e.g. JS-rendered\nprices). State in `reason` that you checked the canonical enumerating page and the\nentity was not present. *Example (FN5): claim \"99.99% SLA tier\" while the reliability\npage lists only 99.9% → absence of any 99.99% tier = `not_grounded`.*\n\n**R3 — Frame/unit replacement (fixes FN4).** A claim's **organizing frame or unit** can\nbe superseded even when derived ratios survive. If the page shows the claim's framing\nhas been **replaced** (e.g. \"1 Unit Capacity\" → \"Quota Tiers\"; a renamed/retired\nmetric), the claim is `not_grounded` even if some embedded numbers still appear\nsomewhere — the claim describes a world that no longer exists. Check that the *unit and\nstructure* the claim assumes still match the current page, not just the digits.\n\n**R4 — Current row, never a legacy row (fixes FN6).** Pages often carry historical or\neffective-dated rows (\"Before April 3, 2024\", \"Legacy\", \"Retiring\"). A claim is\n`grounded` only if it matches the **current/effective** row. Matching a clearly\ntime-stamped *past* row is `not_grounded` (the value has since changed). Always locate\nthe row that applies *today*. *Example (FN6): storage limits matching only the\n\"Before April 3, 2024\" row while current limits differ → `not_grounded`.*\n\n**R8 — Multi-part claims: every load-bearing part must hold (v3.1 — fixes the\n`ai-foundry-dr#9` FN).** A single claim often bundles **several load-bearing\nsub-assertions** (a status AND a region; a capability AND a named target; a date AND a\nGA level). Verify **each load-bearing part separately**. If **any one** load-bearing\npart is contradicted by the source, the whole claim is `not_grounded` — even when the\nother parts check out. Do not let a correct first half earn a `grounded` for a wrong\nsecond half. *Example (v3-FN): \"Global training (Public Preview), cheaper, no data\nresidency; use regional in Norway East\" — the GA-vs-Preview part and the \"no residency\"\npart hold, but **Norway East is a Global (non-residency) training region, not a regional\none** → one load-bearing part is wrong → `not_grounded`.* (R8 is the mirror of R6:\nR6 forgives an **omitted, non-load-bearing** detail; R8 condemns a **stated,\nload-bearing** part that is wrong. Decide first whether the part is load-bearing — if\nthe claim *asserts* it and a reader would act on it, it is.)\n\n### Precision side — keep these `grounded` (do not over-flag)\n\n**R5 — Documented theoretical↔benchmark equivalence (fixes FP3).** Do not flag a\nnumeric claim merely because the exact string is not verbatim, when the asserted value\nis the **documented theoretical or benchmark equivalent** of what the page states and\nboth trace to Microsoft sources (e.g. a theoretical max vs a measured benchmark of the\nsame technique, same order of magnitude, same direction). The exact-value rule targets\n*drifted/contradicting* values — not two Microsoft-sourced expressions of the same\nfact. If the page substantiates the magnitude and the technique, keep `grounded` and\nnote the equivalence in `reason`.\n\n**R6 — Core grounded, detail omitted ≠ ungrounded (fixes FP4).** Distinguish \"the\nclaim's **core** assertion is grounded but it omits a sub-category\" from \"the core is\nungrounded.\" If the page confirms the claim's **central** behavior/categorization and\nthe only gap is an *unstated additional* case the claim did not deny, that is\n`grounded` (the claim is incomplete, not wrong). Reserve `not_grounded` for when the\npage **maps the core differently** or the claim **asserts** something the page\ncontradicts. Omission ≠ contradiction. (Contrast R8: an omitted case is forgiven here;\na *stated* but wrong load-bearing part is not — that is R8's domain.)\n\n**R7 — Follow the capability to its canonical page; don't punish illustrative numbers\n(fixes FP5; v3.1 sharpens the load-bearing carve-out).** If a claim asserts a **real\ncapability** and the cited `evidence_url` does not foreground it, search for the\n**canonical** page that documents the capability before judging — do not return\n`not_grounded` merely because the *cited* page was a weak choice. And when a capability\nis solidly grounded, do **not** flag it over an *illustrative* attached number (e.g.\n\"~0 RTO/RPO\", \"≈15 min\") that the claim offers as an order-of-magnitude illustration\nrather than a cited spec. Judge the **capability**; treat an illustrative figure as\ngrounded if the capability is.\n\n> **⚠️ Load-bearing carve-out (v3.1, fixes the `token-usage#3` FN).** R7's leniency\n> covers only *illustrative* values. It does **NOT** cover a value or **exact string**\n> that **IS the assertion** — a metric name, an API field, an SDK identifier, an enum\n> value, a specific date/version. When the claim's load-bearing content is the literal\n> name/string itself (e.g. \"the metrics are `PromptTokens` and `CompletionTokens`\"),\n> the exact-value rule applies in full: if the live page names them differently\n> (`ProcessedPromptTokens` / `InputTokens` / `GeneratedTokens` / `OutputTokens`), the\n> claim is `not_grounded`. \"Follow to the canonical page\" means find the **right\n> names**, not rescue wrong ones. A reader would copy that string into code; an\n> illustrative magnitude they would not.\n\n---\n\n## Procedure (per claim)\n\n1. **Identify the volatile assertion(s)** in the claim text — and when the claim\n bundles several (R8), enumerate **each load-bearing part**. The `claim_type` tells\n you what to check:\n - `version` → model/API version, GA date, context window, max output, training cutoff\n - `tpm` → tokens-per-minute / throughput / quota numbers\n - `sku` → SKU name, tier, PTU minimums, deployment type\n - `region` → regional availability\n - `status` → GA / preview / retirement / deprecation status\n - `taxonomy` → categorization, capability mapping, which-feature-does-what\n2. **Fetch the cited source** with `microsoft_docs_fetch` on the claim's\n `evidence_url`. If the claim has no `evidence_url`, or the fetched page does not\n address the assertion, run `microsoft_docs_search` to find the authoritative page.\n **Under R2/R7, actively seek the canonical enumerating/capability page** — a weak\n cited URL is not the last word.\n3. **Exact-value entailment check** each checkable value (and each load-bearing part\n under R8), then apply the calibration rules R1R8. Classify which rule(s), if any,\n govern the claim. For R1, first decide whether the bound is a **lower** bound (floor)\n or an **upper** bound (ceiling) — they invert.\n4. **Strict evidence rule:** a `grounded` or `not_grounded` verdict REQUIRES a verbatim\n quote you actually fetched from a `learn.microsoft.com` URL. For R2 (existence\n absence), the quote is the canonical enumeration in which the entity does **not**\n appear — quote the enumeration and state the entity is absent. No quote → `source_silent`.\n\n## Hard rules\n\n- Verify against the fetched page only. Do not rely on prior knowledge of model\n specs / prices — those are exactly what may have drifted.\n- Stable identifiers are not volatile and are not your job to refute: regulation year\n (2024/1689), case numbers (C-311/18), standard version names (OWASP LLM Top 10\n 2025, MADR v3.0), file names. If a claim is purely such an identifier, judge it on\n whatever volatile value it carries, else `source_silent`.\n- One verdict per claim. Return EXACTLY the JSON below — no prose, no markdown fence.\n- `evidence_quote` = the verbatim sentence/value from the fetched page that drove the\n verdict (empty string for `source_silent`). `evidence_url` = the page you actually\n used (may differ from the cited one if you fell back to search).\n- `rule` = which calibration rule governed, if any (`R1``R8`), else empty.\n\n## Batch to judge (from `skills/ms-ai-governance/references/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md`)\n\n[\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md#1\",\n \"claim\": \"Azure Cost Management-verktøy: Cost Analysis, Budgets, Cost Alerts, Advisor Recommendations\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cloud-adoption-framework/ready/azure-setup-guide/manage-costs\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md#2\",\n \"claim\": \"Budsjetter kan opprettes på subscription-nivå, resource group-nivå og tag-basert\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md#3\",\n \"claim\": \"Budgettyper: Actual cost budget (faktisk forbruk) og Forecasted cost budget (forventet forbruk)\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md#4\",\n \"claim\": \"Budget alerts: e-postvarsel ved 50 %, 80 %, 100 % av budsjettgrensen; integrasjon med Azure Logic Apps for automatiske tiltak\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md#5\",\n \"claim\": \"Azure Cost Management støtter tag inheritance og cost allocation rules for automatisk tagging av underliggende ressurser\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/overview-cost-management\"\n },\n {\n \"id\": \"ms-ai-governance/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md#6\",\n \"claim\": \"Kostnadsdata kan eksporteres automatisk til Azure Storage (daglig/ukentlig/månedlig), med detaljert rapportering per ressurs, tag og tidsperiode\",\n \"claim_type\": \"taxonomy\",\n \"evidence_url\": \"https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-improved-exports\"\n }\n]\n\n## Output (strict JSON, no fence)\n\n```\n{\"file\":\"skills/ms-ai-governance/references/norwegian-public-sector-governance/budget-and-accounting-ai-costs.md\",\"results\":[\n {\"id\":\"<claim id>\",\"judge_verdict\":\"grounded|not_grounded|source_silent\",\"rule\":\"<R1-R8 or empty>\",\"evidence_url\":\"<url actually used>\",\"evidence_quote\":\"<verbatim quote or empty>\",\"reason\":\"<one sentence: what the source said vs the claim>\"}\n]}\n```\n"
}
]