ms-ai-architect/docs/r11-pilot-results.md

R11 pilot results — measured, 2026-08-03

The §10 acceptance measurement of docs/r11-tiered-fix-design.md, run against the live ledger. No KB file was edited and no ledger record was written.

Instrument: scripts/kb-eval/lib/fix-op.mjs (+ tests/kb-eval/test-fix-op-classify.test.mjs, 49 tests after §4b) driven by scripts/kb-eval/classify-fix-ops.mjs. The classifier is the O1 driver with writes disabled — it constructs the swap and checks the §4 invariant, so measurements 1 and 3 come out of the mechanism that would later touch the corpus, not out of a proxy heuristic.

Artefact: scripts/kb-eval/data/r11-pilot-classification.json (untracked, regenerable; per-flag records so the run can be re-analysed without re-running). It holds the pilot run — node scripts/kb-eval/classify-fix-ops.mjs --write. Every corpus-wide figure below is from --threshold 1, and the per-table reproduce command is stated where it is used.

Two different 202s. This population is 202 flags. §3's "202 flags whose claim and quote contain a numeric token" is a different 202, measured over the full 712-flag population. They are unrelated.

---

1. The headline

Nine provable, correct value swaps exist in the entire 776-flag not_grounded population — 1.2 %. The machine half of the R11 tiering buys nine edits. Every other flag needs a human.

Population	Files	Flags	O1 admitted	O1 hand-verified correct
Pilot (not_grounded ≥ 7)	24	202	2	2
Whole not_grounded corpus	218	776	15	9

These are numeric-path figures, and they stay that way. §4b (the status synonym table) was implemented afterwards and adds a separate class with its own hand-verification — see §8. A run today prints O1 = 7 (pilot) and 23 (corpus) because the status proposals are included in the total; the numeric line above is unchanged and is still what s4_as_written compares against.

Reproduce: node scripts/kb-eval/classify-fix-ops.mjs (pilot) and node scripts/kb-eval/classify-fix-ops.mjs --threshold 1 (corpus). The nine are enumerated with verdicts in appendix A — that hand-verification is the only thing separating 9 from 15, so it is recorded rather than left in a session transcript.

This is the answer §10 asked for, and it is materially worse than the design assumed: "If the split is materially worse than assumed, that is known after one session rather than after ten."

The O2 half, measured afterwards (§9), does not rescue the number. Of the 46 R8 multi-part claims that have the O2 shape, 17 are candidates and 2 are clean subtractions on the available evidence; all 29 non-candidates fail because the source supplies a corrected value, which makes them swaps or rewrites. O2's value in R11 is triage — it tells a human which 17 to look at first — not automation.

2. §4 as written is not sufficient — measured, not argued

§4 claims its invariant is "deliberately stronger than human review at scale." It is not. Run exactly as specified over the pilot, it admitted 6 swaps, of which 4 were wrong — precision 2/6:

Proposed swap	Why it is wrong
30-dagers → 24-dagers	Unit crossing. The quote says 24 hours.
3000 requests/sekund → 50	Metric crossing. The quote is a query throttle per index; the claim is an indexing rate per replica.
Microsoft Agent 365 → Agent 7	Identifier mutilated. The 7 was harvested out of E7.
text-embedding-ada-002 → ada-2	Identifier mutilated. The 2 came from a dimensions column.

The defect is structural, not incidental. §4 constrains where the new value came from (verbatim in the cited quote) and what the edit looks like (one line, rest byte-identical). It constrains nothing about whether the two tokens denote the same quantity. Same-type-and-provenance is not same-referent.

The added condition

contextCorresponds() requires the token to sit under the same label or the same trailing unit on both sides. It is deliberately lexical, with no translation table: dokumenter is not taught to equal documents, because a synonym table introduces a new fact source and is an operator decision, not an engineering one. Consequence, measured: a swap is provable essentially only where the context is language-neutral — a URL, a code sample, a parameter key.

3. The condition is necessary but still not sufficient

Corpus-wide the context condition cut 38 admissions to 15. Hand-verifying all 15 splits them cleanly by token type:

Reproduce: node scripts/kb-eval/classify-fix-ops.mjs --threshold 1 (the persisted artefact is the pilot run — the corpus-wide tables in §1 and §3 come from this threshold-1 run). Per-proposal verdicts: appendix A.

Token type	Proposals	Correct	Wrong	Unverified	Failure mode
iso_date	9	9	0	0	— every one is an api-version= bump in a URL or code sample
number	5	0	3	2	AI-900 → AI-901, gpt-4o → gpt-5.1o (×2)
version	1	0	1	0	Java agent 3.7.5 → 3.4.0 — a downgrade

A matching identifier prefix (AI-, gpt-) satisfies the context condition while the digit is part of a name, not a quantity. Only iso_date survives hand-verification, and the report marks it as the sole recommended class (o1_recommended). number and version proposals must not be applied.

4. The four §10 measurements

1. O1 / O2 / O3 split. O1 = 2/202 on the pilot (9/776 corpus-wide, safe class only). O2 is undetermined — it does not exist as a class until §5 is ratified, so every non-O1 item is O3 by design. §5 was ratified 2026-08-03, and O2 has since been measured over the class where it can exist at all — see measurement 2 and §9. It remains unmeasured outside R8 ∧ MULTI_PART_CLAIM; the classifier still routes every non-O1 item to O3, so O3 ≥ 200/202 stands as the machine's own partition.

2. How much of R8 resolves as O2. MEASURED 2026-08-03 — see §9. R8 is 87/202 on the pilot (366/776 corpus-wide) and yields zero O1. Of the pilot's 87, 46 are structural enumerations (R8 ∧ MULTI_PART_CLAIM) — the O2 candidate shape. Which of them subtract cleanly turns on the judge's prose reason, and no regex reads prose, so this was done by prose classification: 17 of 46 (37 %) are O2 candidates, 29 are O3. All 29 are foreclosed by condition 3.

3. O1 abort rate: 99 % (200/202). Typed, because "99 %" alone is not actionable:

   Code	Pilot	Class
   MULTI_PART_CLAIM	96 (47.5 %)	intrinsic — not a value swap at all
   MULTI_VALUE_TOKEN	29	intrinsic
   NO_VALUE_TOKEN	28	intrinsic — the claim asserts prose
   STATUS_SYNONYM	15	operator question (§6.2)
   NOT_VERBATIM	15	intrinsic
   LOCATOR_AMBIGUOUS	7	fixable engineering gap
   MULTI_REPLACEMENT	6	intrinsic
   CONTEXT_MISMATCH	4	intrinsic — these are the 4 wrong edits above

   Only 7 of 200 aborts (3.5 %) are a fixable engineering gap. More locator engineering cannot move the O1 number materially.

4. Review throughput per class. NOT MEASURED. It requires human review sessions, which have not happened. Recording it as measured would be false.

5. Two further findings

F1 — subtraction can leave a misleading remainder. §5 argues O2 "cannot introduce a new error, because it asserts strictly less." True of the sentence, false of the reader's inference. Real case: *"Deep Research tool (o3-deep-research

• Bing) er GA (juni 2025)"* where the source says the tool is deprecated. Subtracting er GA (juni 2025) leaves the tool standing in a list of available tools. Strictly less asserted; still misleading. O2 therefore still requires a human to look at the remainder — cheaper than O3 (no fact-finding) but not mechanical.

F2 — subtraction can destroy true information. Real case: a list of seven prebuilt model IDs where the judge found six correct and prebuilt-check wrong — the real ID is prebuilt-check.us. Subtraction drops a model that exists; the correct fix is a swap. Subtraction is not the safe default everywhere.

F3 — disposition carries zero information. It is outdated on 202 of 202 flags. docs/r11-flag-format-2026-07.md specifies not_grounded → {outdated, wrong} with "the human assigns which at R11", but the pass hard-assigned outdated. Do not use it as a classifier signal. Spec/data divergence, recorded.

F4 — claims are not file text. claim is an LLM-extracted, translated restatement: 0 of 202 match their file line verbatim, and 188 share no 40-char run with it. For table claims, line points at the header, not the value. This is why the locator exists at all, and why it searches the enclosing block rather than the line.

6. Operator decisions — ALL THREE RATIFIED 2026-08-03

All three were put to the operator with the recommendations below and all three were accepted as recommended. The contract text now lives in docs/r11-tiered-fix-design.md §4a/§4b/§5; this section records what was asked and what the answer was.

None of the three is implemented yet. The classifier still aborts STATUS_SYNONYM and still routes every non-O1 item to O3. A later session builds against the ratified contract — it must not assume the code already honours it.

1. Ratify O2 (§5)? → RATIFIED, with the remainder check (not as a blanket rule), exactly as F1/F2 above argued. Contract: design doc §5, three conditions, human-confirmed.
2. Amend §4 with a ratified synonym table? → RATIFIED, narrow and closed. Contract: design doc §4b — four label rows, closed table, complete-label-only, corpus-side value with the file's own markup preserved. Unlocks up to 54 corpus-wide flags.
3. Is O1 worth building at all? → KEPT, locked to iso_date. Contract: design doc §4a condition 5 — a driver may apply iso_date proposals and must never apply number or version ones. Nine edits corpus-wide.

7. What this does not change

The design's core reading survives: the expensive half (locating the source, reading it, extracting the deciding passage) was already paid for by the judge pass, and §7's freshness re-fetch per distinct URL is untouched. What the pilot falsifies is the assumption that a meaningful share of that evidence converts into machine-provable edits. It does not. R11 is a human review programme with a nine-item machine assist, and its leverage lies entirely in the O2 decision.

---

8. §4b implemented — the status class measured, 2026-08-03

The ratified synonym table (docs/r11-tiered-fix-design.md §4b) is implemented in fix-op.mjs and the class is measured. Reproduce: node scripts/kb-eval/classify-fix-ops.mjs --threshold 1 (status_synonym block).

Population	STATUS_SYNONYM flags	Proven by §4b	Still aborting
Pilot (not_grounded ≥ 7)	15	5	10
Whole not_grounded corpus	54	8	46

Why the other 46 abort, corpus-wide — this sub-distribution is the actionable part, because the top-level STATUS_SYNONYM count alone says nothing:

Reason	N	What it means
NO_COMPLETE_FILE_LABEL	27	the file writes the status inside a sentence — (preview) in a list item, **DSPM (preview):**, "[Preview]: …" in a JSON string. Constraint 2 refuses these, correctly.
NO_SOURCE_STATUS	15	the cited quote carries no listed lifecycle phrasing at all — the flag was never a status swap.
SOURCE_STATUS_AMBIGUOUS	2	the quote asserts two different rows (e.g. "…is now generally available. Partner solutions remain in preview.").
FILE_ALREADY_MATCHES	2	file and source agree; the mismatch was in the LLM-extracted claim, not in the corpus.

The class is REVIEW-grade, not apply-grade — 5 of 8 correct. All eight were hand-judged against the cited source (appendix B). Three defects, all one family: §4b binds the table, the completeness of the file label and the written value, and nothing about whether the source phrasing refers to the row's own subject. That is the same provenance-without-referent defect that falsified §4 (§2), now reproduced in the status class. status is therefore deliberately absent from o1_recommended: the machine writes nothing, and every proposal reaches a human.

Two candidate conditions, costed over the eight

Neither is implemented — extending a table the operator ratified as closed is an operator decision, exactly as condition 5 was in §4a. Both are pure gain on this population (they kill wrong proposals and no correct one), which is the number the decision needs:

Candidate	Kills	Correct proposals lost
A — count a bare GA in the quote as a GA-row source phrasing, so a quote saying both GA and "public preview" becomes ambiguous	1 (onelake:198)	0
B — abort when the quote is a multi-entity enumeration (≥ 2 pipes, or a numbered list)	2 (onelake:198, owasp:79)	0

B subsumes A on these eight. Neither catches security-copilot-integration.md:94, where the quote is prose and the (Preview) marker simply belongs to a different agent. A stricter referent-name condition (the row's subject must appear in the quote) would catch it — and would also kill two correct proposals (:83, :93), where the source names the capability rather than the agent. That trade is real and is why this is put to the operator rather than shipped.

9. §10 measurement #2 — how much of R8 resolves as O2, measured 2026-08-03

The last machine-answerable pilot measurement. §4.2 recorded it as "not answered, and not answerable by machine" — true of a regex, not of prose classification, which is what this ran. Reproduce the verification and the tally: node scripts/kb-eval/check-o2-returns.mjs.

Population. The 46 pilot flags that are R8 ∧ MULTI_PART_CLAIM — the O2 candidate shape, re-derived from the ledger, not read from a plan (classify-fix-ops.mjs --threshold 7). 17 distinct files.

Method. Eight subagents, six items each, classifying against §5's three conditions. Constraints, all deliberate: read-only (no writes, no commits); no web or MCP lookups — O2 is defined by requiring no new fact-finding, so the evidence_quote is the only source evidence a classifier may use; and every proposal must be obtainable from the file text by deleting characters only. Because claim matches its file line verbatim in 0 of 202 cases (F4), each classifier had to open the actual file and locate the real text rather than edit the restatement. All 46 located it; locator_failed is 0.

Verdict	N	Share
O2 candidate	17	37 %
O3	29	63 %

The result that matters is not the split — it is what blocks the other 29.

Blocking condition(s)	N
condition 2 + condition 3	21
condition 3 alone	3
all three	5

All 29 are foreclosed by condition 3: the source supplies a corrected value, so the fix is a swap or a rewrite and subtraction would destroy true information. Condition 1 — "asserts strictly less", the one that sounds like the hard one — blocks only 5, and never alone. This is F2 (§5) reproduced at scale: prebuilt-check.us, prebuilt-mortgage.us.closingDisclosure, Set-DlpCompliancePolicy, jensen_shannon_distance, F300 = 384 GB, DurationMs / ResultSignature, Claude 4.5 → 4.6. R8's failing multi-part claims are predominantly a wrong-value class, not a surplus-specificity class. The design's reading of R8 in §3 — "where the grounded part stands on its own, the fix is O2" — holds for a minority of the class.

Machine verification of the returns (V1/V2/V2b, §9.1). 46 of 46 pass V1: the quoted file text occurs verbatim in the named file, æ/ø/å and markup intact. 16 of the 17 O2 proposals are deletion-only; one is flagged (feedback-loops-continuous-improvement.md:555, where Automatically add → Add recapitalises rather than merely deletes). That is a text change, not a subtraction, and it goes to a human as such.

The 17 are candidates, not admitted edits — and the split inside them is the honest number:

	N
conditions 2 and 3 both affirmatively yes	2 (idx 8, 17)
at least one condition marked human_must_confirm	15
classifier confidence high	1

So the machine's own reading is that 2 of 46 (4 %) are clean subtractions on the evidence available, and 15 more are worth a human's time. This is a triage, not a machine assist. It is not measurement #4 — review throughput still requires human review sessions that have not happened (§4.4) — but it is the input #4 needs: it says how many items enter review and in what state, which is the half of throughput that does not require a stopwatch. The recurring reason for human_must_confirm on condition 3 is structural and worth recording: the removed material is often true of something else (Purview really does classify data; the Communication Compliance template really exists; Redis really is in Norway West) — it is merely false of the subject the row names. Deleting it is defensible; relocating it may be better. That is a judgement about the corpus, not about the source, which is exactly why §5 put conditions 2 and 3 in human hands.

Extrapolation, flagged as such. MULTI_PART_CLAIM is 161 corpus-wide under R8. At the pilot's 37 % that is ~60 O2 candidates and ~6 clean ones. This is an extrapolation from one measured sample, not a measurement, and the pilot was deliberately drawn from the densest files.

9.1 What the machine checks, and what it deliberately does not

scripts/kb-eval/lib/o2-return-check.mjs (25 tests). The checks do not decide O2 — conditions 2 and 3 stay human by ratified contract. They bound the two failure modes a human reviewing 46 proposals cannot catch cheaply:

• V1 — the quoted file text must occur verbatim in the file. Catches invented text and silent æ/ø/å transliteration. Applied to every row, not just the O2 ones: an O3 verdict resting on invented text is equally wrong, merely wrong in the safe direction.
• V2 — the remainder must be obtainable by deleting characters only.
• V2b — word-level and case-sensitive, because V2 alone is too weak: deleting a leading word and recapitalising the next passes the character test, since the capital already existed inside the deleted word. This check was added after wave 1 produced exactly that case.
• V3 — schema completeness and verdict/condition coherence.

The raw returns are committed under scripts/kb-eval/data/r11-o2-returns/ — they are evidence, not regenerable output, same discipline as appendices A and B.

9.2 The two affirmative candidates, hand-verified — one of them fails

Before putting anything in front of a human ratifier, the two candidates the classifier marked affirmative on both human conditions (idx 8, 17) were checked by hand, 2026-08-03. The check was not a re-reading of the source: it was a check of the classifier's own cond-2 reasoning, which cites other lines in the same file as its justification. V1/V2/V2b never touch those citations — they bound the quoted block and the remainder string, nothing else.

Anchoring, verified for all 17. Each file_text_verbatim occurs exactly once in its file (17/17), so text-anchored editing is unambiguous. The line numbers are not: line differs from real_line in 9 of 17 records. Any edit must be anchored on the verbatim block, never on the line number.

idx 17 — holds. Both cross-references check out: line 164 does carry Start med 0.15, and the policy sample does use score-threshold="0.15" — at line 238, not 239 as the classifier wrote (an off-by-one in the citation; the substance stands). The three bands appear nowhere else in the file (grep for the band values and their labels returns only lines 254-256, the block itself), so deleting them leaves nothing dangling internally. Two cosmetic residues for the ratifier, both already visible in the record: the heading keeps a now-trailing colon, and the **Verified** (Microsoft Learn - Enable semantic caching for LLM APIs) stamp on line 258 afterwards stamps only the direction statement. Neither makes the remainder false.

idx 8 — not ratifiable as written, under either reading of cond 2 — but the reason and the remedy differ, and choosing between them is an operator call.

First, the fork, because everything below depends on it. Cond 2 says the remainder must not be misleading. Its scope was never fixed:

• Broad reading — misleading to a reader of the file. Then the rest of the file is in scope, and a remainder that contradicts a passage seventy lines down fails.
• Narrow reading — misleading as a statement of what the source grounds. Then only the edited passage is in scope, and a contradiction elsewhere in the file is a separate ungrounded claim, to be flagged on its own, not a defeater of this subtraction.

The operator ratified the broad reading, 2026-08-03. Cond 2 is measured against the whole file: a remainder that contradicts the file it sits in is misleading, whatever the source says. The stated ground is that these files are publicly distributed and read as wholes — a self-contradicting file is a trust defect regardless of which half is wrong. The narrow reading (cond 2 scoped to the edited passage, with contradictions elsewhere handled as separate ungrounded claims) was considered and rejected. Recorded here because the fork was real and a later run must not silently re-open it.

The classifier justified cond 2 on one of the two sub-deletions and never checked the other:

• Sub-deletion (b), the Optional: Application Insights bullet: verified. Lines 202-203 do cover Application Insights on their own terms (**Azure Monitor + Application Insights** (Verified) / "Drift metrics emitteres til Application Insights"), so removing the prerequisite bullet leaves no false implication.
• Sub-deletion (a), eller managed compute cluster: fails. The same file documents **Managed Compute Cluster** (for store volumer) as a real compute option at lines 263-265, with pricing and a usage recommendation. Deleting the alternative from the prerequisites leaves the file asserting serverless Spark as the only compute requirement seventy lines above a cost section that prices the alternative. That is a misleading remainder under the ratified reading.

The candidate therefore does not go in as written: (a) must be dropped from the subtraction, leaving (b) alone.

There is a further limit the O2 envelope cannot resolve: cond 3 passes for (a) only because the source does not mention managed compute — not-mentioned passes cond 3 by construction. Whether Azure ML actually permits managed compute for monitoring is a fact question, and O2 is defined by zero fact-gathering. It is an operator call, not a gap to be read harder.

A reduced subtraction — sub-deletion (b) alone — would still be deletion-only and is not defeated by anything found here. But it is a different remainder string than the one V2b attested, so it is not machine-clean until check-o2-returns.mjs is re-run against it. The same rule governs any operator amendment, including dropping idx 17's trailing colon: amended remainder → re-run the check before it counts as verified.

The generalisable finding. The classifier judged cond 2 against the source and against citations it chose itself. It did not systematically judge it against the rest of the same file. Under the ratified reading that is a defect: internal consistency is a cond-2 dimension the wave prompts never assigned, and idx 8 — the single high confidence record in the set — is the proof that it bites. The remaining 15 have not had the check. It must be run per candidate before any of them reaches a ratifier, and its output corrects the existing cond-2 verdicts rather than merely adding to them; expect it to move some.

Score after hand-verification: 1 of 46 clears both conditions, not 2. Only idx 17 survives. That is the number §10 measurement #2 should be read with — the classifier's own "2 of 46" counted idx 8 on a cond-2 justification that was half-unchecked.

9.3 The whole-file check run on the remaining 15

Run 2026-08-03, one candidate at a time, over the 15 O2 candidates other than idx 8 and 17 — ten distinct files. No KB file was edited.

Method, and its one calibration. For each record, the deletion segments were recovered by diffing file_text_verbatim against proposed_remainder (word-level LCS), content phrases were extracted from each segment (markdown stripped, Norwegian and English stopwords dropped, contiguous content runs of 2-3 words kept as noun-phrase units), and each phrase was matched case-insensitively against every line of the file outside the verbatim block. The extractor was calibrated on idx 8 before the sweep: it must surface Managed Compute Cluster at line 263 from the deleted eller managed compute cluster. It does, as the top-ranked multi-word hit. Without that calibration a narrower extractor would have returned a clean bill on idx 8 — and silently on others.

The grep is necessary and not sufficient. Two of the findings below have no lexical overlap with the deleted tokens at all. Idx 14's surviving claim is the Norwegian automatisk restating a deleted English Automatically; no token search finds it. Every candidate was therefore also read in place — the enclosing section around real_line, plus every hit line with context — and asked the second question the grep cannot: does the surviving text now claim something broader or narrower than before, and does anything else in the file depend on the version that was there?

Result: 9 clean · 4 contradicted · 2 operator calls.

idx	file:line	outcome	the line that decides it
7	document-intelligence-prebuilt-models.md:79	clean	prebuilt-document occurs only in the deleted row; the one General hit (192) is generalisering; no count binds the table
9	data-drift-monitoring-detection.md:218	clean	the only other Foundry/RAG reference (320) asserts exactly groundedness, relevance — the remainder — and never claims drift detection over grounding data
14	feedback-loops-continuous-improvement.md:555	contradicted (partial)	566: Reviewed documents automatisk tilgjengelige i "Feedback loop" data source når modellen retraines
18	rag-caching-optimization.md:29	contradicted	303-318: a whole section ### Azure AI Search - Built-in Caching, plus 510: Azure AI Search caching | **Verified**
19	rag-caching-optimization.md:297	clean	the deleted bullet is the file's only indexing statement; the code sample's silence about explicit vector indexes predates the deletion
26	transparency-documentation-standards.md:117	contradicted (partial)	300: | **Risk assessment** | Responsible AI Scorecard: Error analysis, fairness assessment |
27	transparency-documentation-standards.md:426	contradicted (partial)	216: - **Copilot Studio**: "Powered by AI" disclosure i chat interface
28	transparency-documentation-standards.md:83	clean	Hugging occurs only in the deleted bullet; no model-card template is claimed anywhere else
31	ai-incident-response-procedures.md:139	clean	legalHold occurs only in this block; enabled at 148 is Blob versioning; 82's "Legal hold på alle artifacts" is consistent with a tags-only hold
33	ai-threat-modeling-stride.md:211	operator call	357 restates both deleted capabilities — but about the CAF document, not about Defender AISPM
36	ai-threat-modeling-stride.md:38	operator call	310: Backdoored models og data poisoning er Critical-severity trusler — unqualified, while the remainder narrows the register to (targeted)
38	data-leakage-prevention-ai.md:396	clean	both deleted policy templates occur only here; the following one-click-policy block names neither
40	supply-chain-security-ai-models.md:133	clean	the CVSS band definition occurs only here; 158's critical vulnerabilities is container-image scanning, a different tool
42	supply-chain-security-ai-models.md:200	clean (strengthened)	every other HuggingFace reference (32, 240, 492, 513) treats it as an unverified source; the deleted bullet was the outlier
45	semantic-caching-patterns.md:436	clean (strengthened)	Norway West appears nowhere else; 303, 451, 488 and 628 are all Norway East

The four contradictions are one shape. In each, an enumerating passage would lose a member that the file continues to assert elsewhere — a section (18), a mapping row (26), a bullet in a sibling section (27), a prose restatement (14). That is idx 8's shape exactly: a list narrowed against a file that documents what was removed. Idx 18 is the hardest of them, because the surviving claim is not a stray sentence but a titled section and a row in the verification table stamping it **Verified**. No deletion confined to line 29 can fix that file; the correct edit is larger than the O2 envelope permits.

Three of the four admit a reduced subtraction, on the same terms as idx 8:

• idx 14 — drop the Automatically→Add half, keep / SharePoint (which is the file's only occurrence). This also removes the V2b machine flag, since the recapitalisation was the flagged part.
• idx 26 — drop item 4 (Error analysis), keep item 5 (Counterfactual analysis); 324, 475 and 693 attach counterfactuals to the dashboard and to GDPR, never to the scorecard. The renumbering artifact (1,2,3,4,6,7) survives either way.
• idx 27 — drop the Chat-interface row, keep the Plugin-actions row; grep -niE "confirmation|plugin" returns line 428 alone.

Every one of these is a different remainder string than the one V2b attested, so none is machine-clean until check-o2-returns.mjs is re-run against it. Idx 18 has no reduction: it is a single deletion.

The two operator calls are a distinct class, and are not being called contradictions. In both, the deleted content survives elsewhere in the file without the remainder becoming false:

• idx 33 — line 357 does say AI asset inventory via Azure Resource Graph and Microsoft Purview Insider Risk Management for prompt-basert data exfiltration-deteksjon, stamped *(Verified MCP 2026-04)*. But it says it about what the Cloud Adoption Framework document now covers, while the deleted bullets attributed those capabilities to Defender for Cloud AISPM. Different subjects, so no contradiction — but the edit's benefit is smaller than it looks, because the content it removes stays in the file under another attribution.
• idx 36 — the remainder narrows the severity register to Data Poisoning (targeted) while line 310 still justifies a recommendation with the unqualified data poisoning er Critical-severity. A narrower statement does not contradict a broader one; it is subsumed by it. What the edit produces is a file whose severity table is more precise than the prose that cites it. Whether that is acceptable, or whether 310 needs the same qualifier, is a judgement about the file — and a companion edit at 310 exceeds the single-locator O2 envelope.

What this does to the score. The sweep settles the whole-file dimension of cond 2 for all 15: affirmative for 9, negative for 4, operator for 2. It settles nothing about cond 3, which stands at human_must_confirm for ten of them. Only one candidate moves into the both-conditions-affirmative class: idx 19, whose cond-2 doubt was itself whole-file-shaped (the classifier worried that the preceding code sample's silence about explicit vector indexes might read as "no setup needed") and whose cond 3 the classifier already marked yes. The silence predates the deletion, and the deletion removes an affirmative false claim, so the remainder asserts nothing the file denies.

Verified score: 2 of 46 — idx 17 and idx 19. Up from 1, and by a different route than the classifier's original 2: idx 8 left the class and idx 19 entered it.

What the sweep says about the method. The classifier's cond-2 column was wrong in one direction only. Of the eleven candidates it marked cond 2 = yes, the whole-file check overturns two (27, 36) and confirms nine. Of the four it marked human_must_confirm, the check clears one (19) and confirms the doubt on three (14, 18, 26) — in every one of those three the classifier had already written the contradicting line number into its own evidence field without treating it as a defeater. The information was in the returns; the contract just never asked the classifier to act on it. That is a prompt gap, not a model failure, and it is cheap to close in a later wave: name internal consistency as a cond-2 dimension and require the citation.

Appendix A — the 15 admitted proposals, hand-verified

Every proposal the classifier (§4 + context condition) admitted over the whole not_grounded population, with the verdict that produced §3's table. A later run that admits a 16th can diff against this list; without it, "9 of 15" is an unreproducible claim.

#	File:line	Swap	Type	Verdict
1	agent-orchestration/agent-evaluation-testing-frameworks.md:56	4.1 → 5 (gpt-4.1-mini → gpt-5-mini)	number	unverified — model identifier; the result is a real model name, but not checked against the source. Not applied.
2	api-management/logging-analytics-ai-traffic.md:49	2023-09-01 → 2025-09-01	iso_date	correct — ARM loggers@ api-version bump
3	azure-ai-services/translator-document-translation.md:162	40 → 10 (MB)	number	unverified — matched on the unit MB, but sync/async limits differ; metric-crossing risk. Not applied.
4	monitoring-observability/log-analytics-kql-ai-queries.md:617	2025-09-01 → 2026-04-01	iso_date	correct — api-version= inside a KQL string literal
5	responsible-ai/responsible-ai-training-awareness.md:77	900 → 901 (AI-900 → AI-901)	number	wrong — certification identifier mutilated
6	bcdr/cost-analysis-dr-configurations.md:120	4 → 5.1 (GPT-4o → GPT-5.1o)	number	wrong — model identifier mutilated
7	bcdr/multi-region-azure-openai-deployment.md:316	2024-06-01 → 2024-10-01	iso_date	correct — api-version= in a management URL
8	ai-security-engineering/ai-prompt-shield-network.md:309	2024-09-01 → 2024-09-15	iso_date	correct — Content Safety api-version
9	ai-security-engineering/content-safety-filter-calibration.md:277	2024-10-01 → 2024-10-21	iso_date	correct — Azure OpenAI api-version in a curl sample
10	ai-security-engineering/jailbreak-prevention-production.md:305	2024-09-01 → 2024-09-15	iso_date	correct — Content Safety api-version in a curl sample
11	cost-optimization/observability-cost-reduction.md:114	3.7.5 → 3.4.0 (Java Agent)	version	wrong — a downgrade; the quote's version is not the claim's referent
12	cost-optimization/vector-storage-cost-optimization.md:266	2025-09-01 → 2026-04-01	iso_date	correct — AI Search api-version
13	cost-optimization/vector-storage-cost-optimization.md:318	2024-02-01 → 2024-10-21	iso_date	correct — embeddings api-version
14	performance-scalability/response-chunking-strategies.md:56	4 → 5.1 (gpt-4o → gpt-5.1o)	number	wrong — model identifier mutilated
15	performance-scalability/token-per-second-optimization.md:295	2024-12-01 → 2025-01-01	iso_date	correct — Azure OpenAI api-version

9 correct · 4 wrong · 2 unverified. All nine correct are iso_date; every wrong one is a digit inside a product, model or certification identifier, where a matching prefix (AI-, gpt-, Agent ) satisfies the context condition while the digit is part of a name rather than a quantity. The two unverified are also number and are excluded by the same class rule — verifying them costs a source fetch each and would move the total to at most 11.

Appendix B — the 8 §4b status proposals, hand-verified

Every status proposal the classifier admits over the whole not_grounded population, judged against the cited source. Same discipline as appendix A: a later run that admits a ninth can diff against this list, and "5 of 8" is otherwise an unreproducible claim.

Reproduce: node scripts/kb-eval/classify-fix-ops.mjs --threshold 1 --write, then read the proposal.type === 'status' items in scripts/kb-eval/data/r11-pilot-classification.json.

#	File:line	Swap	Verdict
1	agent-orchestration/foundry-agent-service-ga.md:68	**Preview** → **GA**	correct — quote: "hosted agents are generally available"; the row's subject is Hosted agents
2	agent-orchestration/foundry-agent-service-ga.md:72	**GA** → **Preview**	correct — quote: "Trigger an agent by using Logic Apps (preview)"; the row's subject is the Logic Apps trigger
3	ai-security-engineering/security-copilot-integration.md:83	Public Preview → GA	correct — quote: "Email and collaboration alert triage capabilities are already generally available (GA)"; the row is the phishing/email triage agent
4	ai-security-engineering/security-copilot-integration.md:93	GA → Preview	correct — quote is from the agent's own doc page: "This feature is in public preview"
5	ai-security-engineering/entra-agent-id-zero-trust.md:439	Public Preview → GA	correct — quote: "The Microsoft Entra Agent ID platform is now generally available"; the row's subject is Entra Agent ID (kjerne)
6	ai-security-engineering/security-copilot-integration.md:94	GA → Preview	unproven — the quote's (Preview) marker belongs to Identity Risk Management Agent, not to Access Review Agent. The judge's prose reason does support preview from a what's-new post, so the outcome is plausibly right; the cited evidence does not establish it. Not applied.
7	data-engineering/onelake-data-strategy.md:198	GA → Preview	wrong — the quote says Lakehouse | Yes | GA. "Public preview" in the same quote belongs to Eventhouse. Killed by candidate A and B.
8	ai-security-engineering/owasp-llm-top10-azure-mitigations.md:79	GA → Preview	wrong — the source marks only Response Completeness as preview; the row covers the groundedness/completeness pair, so the edit makes the groundedness half false. The correct fix is to split the row (O2/O3). Killed by candidate B.

5 correct · 1 unproven · 2 wrong. All five correct ones carry the source phrasing on the row's own subject; all three defects are the referent gap described in §8. No proposal was applied — §4b output is a human review list.

Appendix C — the 17 O2 candidates

Every proposal the prose classification admitted over the 46, with the two human-judged conditions as the classifier left them. Same discipline as appendices A and B: without this list, "17 of 46" is an unreproducible claim. confirm = the classifier marked the condition human_must_confirm, i.e. it could not settle it on the evidence available and is handing it over — not a defect, it is the contract.

The full records, including each proposal's verbatim file text and the exact remainder, are in scripts/kb-eval/data/r11-o2-returns/.

#	File:line	Failing sub-assertion	Cond 2	Cond 3	Confidence
7	ms-ai-engineering/azure-ai-services/document-intelligence-prebuilt-models.md:79	The third table row presenting prebuilt-document (General Document) as a current basic model — the source states the general document model is no l…	yes	confirm	medium
8	ms-ai-engineering/mlops-genaiops/data-drift-monitoring-detection.md:191	Two sub-assertions: (a) eller managed compute cluster as an alternative compute option — the how-to page and the monitor schema require a Spark poo…	yes	yes	high
9	ms-ai-engineering/mlops-genaiops/data-drift-monitoring-detection.md:218	The second sentence, Støtter også drift detection for grounding data i RAG scenarios. — the canonical observability page lists only Evaluation, Mon…	yes	confirm	medium
14	ms-ai-engineering/mlops-genaiops/feedback-loops-continuous-improvement.md:555	Two sub-assertions: 'SharePoint' as a feedback storage service, and the word 'Automatically' in 'Automatically add reviewed samples to training set' …	confirm	yes	medium
17	ms-ai-engineering/rag-architecture/rag-caching-optimization.md:254	The three-band rubric (0.1-0.2 strict / 0.3-0.5 balanced / 0.6-0.8 liberal) — undocumented, and the two upper bands contradict the source's warning t…	yes	yes	medium
18	ms-ai-engineering/rag-architecture/rag-caching-optimization.md:29	The list item 'Azure AI Search (built-in caching av search results)' — the source states each query operates on the current index view with no cachin…	confirm	yes	medium
19	ms-ai-engineering/rag-architecture/rag-caching-optimization.md:297	The bullet "Automatic indexing av vectors" — the judge states vector indexes must be declared explicitly in the indexing policy (only at container cr…	confirm	yes	medium
26	ms-ai-governance/responsible-ai/transparency-documentation-standards.md:117	Items 4 (Error analysis) and 5 (Counterfactual analysis) are listed as Responsible AI Scorecard components, but the canonical scorecard segment enume…	confirm	yes	medium
27	ms-ai-governance/responsible-ai/transparency-documentation-standards.md:426	The 'Chat interface' row (a "Powered by AI" badge in the chat window) and the 'Plugin actions' row (confirmation prompts before sensitive actions) ar…	yes	confirm	medium
28	ms-ai-governance/responsible-ai/transparency-documentation-standards.md:83	The second and third bullets — that Hugging Face model cards are synchronised automatically, and that a template exists for generating model cards fo…	yes	confirm	medium
31	ms-ai-security/ai-security-engineering/ai-incident-response-procedures.md:139	The legalHold object is given a field named "enabled"; the Storage API's LegalHold model exposes tags and hasLegalHold, so the literal field name "en…	yes	confirm	medium
33	ms-ai-security/ai-security-engineering/ai-threat-modeling-stride.md:211	Two parts attributed to Defender for Cloud AI Security Posture Management that the AISPM page does not support: the discovery mechanism "(via Azure R…	yes	confirm	medium
36	ms-ai-security/ai-security-engineering/ai-threat-modeling-stride.md:38	The "/indiscriminate" qualifier, which extends the Tampering placement and the Critical severity to indiscriminate data poisoning; the source gives t…	yes	confirm	medium
38	ms-ai-security/ai-security-engineering/data-leakage-prevention-ai.md:396	The listing of 'DSPM for AI - Unethical behavior in AI apps' and 'DSPM for AI - Protect sensitive data from Copilot processing' as Insider Risk Manag…	yes	confirm	medium
40	ms-ai-security/ai-security-engineering/supply-chain-security-ai-models.md:133	The third bullet 'CVE severity mapping' presented as a category of alert that dependency scanning generates; severity is a property of an alert, …	yes	confirm	medium
42	ms-ai-security/ai-security-engineering/supply-chain-security-ai-models.md:200	The second bullet presenting the HuggingFace Registry as a Microsoft channel for verified models with provenance tracking; the source calls it a comm…	yes	confirm	medium
45	ms-ai-security/cost-optimization/semantic-caching-patterns.md:436	The '/West' half of the region pair, i.e. the standing implication that Azure OpenAI can be deployed in Norway West.	yes	confirm	medium

2 affirmative on both conditions (8, 17) · 15 needing a human call. Item 14 additionally carries a machine flag: its remainder recapitalises rather than deletes (§9.1, V2b), so it is a text change and must be reviewed as one.

⚠️ The cond 2 column above is the classifier's claim, not a verified fact — and it has now been corrected. All 17 rows have had the whole-file check (idx 8 and 17 in §9.2, the other 15 in §9.3). Read the column together with §9.3's table, which overrides it:

• Contradicted, not ratifiable as written: idx 8, 14, 18, 26, 27. All but 18 admit a reduced subtraction; every reduction is a new remainder string and must be re-run through check-o2-returns.mjs before it counts as verified.
• Operator call: idx 33, 36 — deleted content survives elsewhere in the file without the remainder becoming false.
• Clean on the whole-file dimension: idx 7, 9, 17, 19, 28, 31, 38, 40, 42, 45. Cond 3 is still human_must_confirm for most of them; clean here means cond 2 only.

The verified score is 2 of 46 — idx 17 and idx 19, not the classifier's original 2 (idx 8 left the class, idx 19 entered it).

Reproduce the tally and the machine checks: node scripts/kb-eval/check-o2-returns.mjs.