fix(ms-ai-architect): G7 idx-26j + idx-26k lukket — footerens referent var alt blandet, av en ratifisert edit

Én form-beslutning, to filer. idx-26k forbyr eksplisitt å lukkes alene, så
enheten var formen, ikke fila.

Det som gjorde referenten avgjørbar:
- 99675d5 (idx-26h, samme dag) satte «Unique sources 15 -> 17» og lot «5» og
  «80 %» staa. Blokken hadde alt blandet referent — mikset av programmet selv.
- INGEN generasjonslogg finnes i repoet. Det fjerner «5 -> 6» fra
  mulighetsrommet under BEGGE lesninger, ikke bare den ene.

Korpusmaaling, haandklassifisert FOER tallet gikk inn i ratifiseringslabelen:
23 footerlinjer i 23 filer — 10 konsistente, 8 inkonsistente, 1 tvetydig,
4 ikke sjekkbare. Regex sa 11; tre var artefakter. Alle 8 avvik gaar SAMME
vei (oppgitt < sum), null motsatt — evidens for at tallet aldri ble
transkribert fra en kjoering.

Anvendt: «Total MCP calls» slettet i begge filer. «Unique sources: 17 URLs»
beholdt (verifisert). «80/20» -> strukturell telling med definert nevner.
Header -> «GA / Preview (Responsible AI scorecard)» ved presedens, ikke
gjenaapnet; parentesens FULLSTENDIGHET verifisert (7 preview-treff gjennomgaatt).
Linje 101 «model-» -> «modell-».

ERSTATNINGEN FORFATTET DEFEKTEN, OG BLE TATT FOER COMMIT: foerste utkast
«12 av 17 kilder verifisert via MCP (nr. 1-12)» ville paastaatt at kilde #7 er
MCP-verifisert — #7 er selv merket «(Status: Baseline …)». Skipet ordlyd teller
bolk-medlemskap alene og paastaar ingenting om enkeltkilder.

Bokfoert, ikke reparert:
- idx-26s: kilde-partisjonen motsier seg selv to steder, motsatt vei (#7, #17).
  Ellevte gang defekten ligger VED SIDEN AV editen. Tellingen 12/5 er invariant
  under begge losninger, saa den skipede linja overlever uansett utfall.
- idx-26l..idx-26r: én entry per beroert fil (6 inkonsistente + 1 tvetydig),
  hver med ordrett anker verifisert unikt. En klasse-bred entry ville latt 20
  filer vaere usynlige for check-g7-queue.mjs — 795d494-feilen i klasseskala.

Ingen re-datering kreves noe sted, og det ble ADJUDISERT foer skriving:
stemplene slutter paa linje 378, footeren staar under intet stempel, og
stempelet paa 131 er alt datert 2026-08-09 og hevder kildeverifisering, ikke
ortografi.

Koe: 14 aapne / 13 resolved. Suite: 1047/1047.

docs/r11-pilot-results.md

@@ -1584,3 +1584,113 @@ have triggered.
   were not re-fetched.
 
 **Queue:** 8 open / 11 resolved. **Suite:** 1047/1047.
+
+---
+
+### 9.14 idx-26j and idx-26k closed — the footer's referent was already mixed, by a ratified edit
+
+Two entries, one form decision. idx-26k exists only because idx-26j predicted the defect would
+recur, and idx-26k's own text forbids closing it alone: *"idx-26j owns that decision and this
+entry is its second measured instance."* So the unit of work was the form, not the file.
+
+#### What made the referent decidable
+
+The entry framed it as a choice between two readings of the footer provenance block — a record
+of the original generation run, or a description of the file's current provenance — and said no
+reader and no check could adjudicate it. Two measurements collapsed that.
+
+**First: the block already carries a mixed referent, and a ratified G7 edit mixed it.** `99675d5`
+(idx-26h, the same day) changed `**Unique sources:** 15 URLs` → `17 URLs` while leaving
+`**Total MCP calls:** 5` and `**Confidence:** 80%` untouched. One field was maintained as current
+provenance; two were left frozen. The precedent inside the block was therefore already set, and
+set by the programme itself.
+
+**Second: no generation log exists.** Grepped `scripts/` and `docs/` for any artifact recording
+MCP calls per file — only the queue and this document mention them, both as analysis rather than
+as a run record. That matters more than which reading wins: it removes `5 → 6` from the option
+space under **both** branches, not just one. A number that cannot be checked under the historical
+reading either is not a historical fact, it is an unsourced claim wearing a date's clothes.
+
+#### The corpus measurement — and a count corrected before it was used
+
+23 footer lines in 23 files state a total followed by a per-tool enumeration. A regex pass
+returned 11 inconsistent. Hand-classification returned **8**:
+
+| | count | |
+|---|---|---|
+| internally consistent | 10 | components sum to the stated total |
+| **internally inconsistent** | **8** | stated total ≠ sum |
+| ambiguous | 1 | `foundry-agent-service-ga.md:560` — reading *is* the question |
+| not checkable | 4 | no per-tool numbers to sum |
+
+The three regex artifacts were `translator-document-translation.md:402` and
+`rag-cost-optimization.md:558`, which spell the arithmetic out (`4 (docs_search) + 3 (docs_fetch)
+= **7**`) and are correct, and `foundry-agent-service-ga.md:560`, whose `4` inside a descriptive
+clause the pattern read as a component. This is the §9.13 lesson applied one step earlier: the
+number was going into an operator-facing ratification label, so it was hand-verified **before**
+the question was asked rather than after the answer came back.
+
+**All eight inconsistencies run the same direction** — stated total *below* the sum, never above.
+A transcription error would scatter. A one-sided bias is evidence the total was never transcribed
+from a run at all.
+
+#### Ratified and applied
+
+Current provenance, one referent for the whole block. `**Total MCP calls:**` **deleted** in both
+files — unmeasurable, with no artifact that could ever adjudicate it. `**Unique sources:** 17
+URLs` **kept**, verified true (17 numbered entries, 17 distinct URLs). The uncheckable
+`80% Verified (MCP), 20% Baseline` became a structural count with a defined denominator. The
+sibling needed only the deletion: its `**Total kilder**: 16 (9 verified fra MCP, 7
+baseline/code samples)` was set by idx-26g and re-verified here by counting, and its
+`**Confidence vurdering**` is a banded per-topic list, not a percentage split.
+
+Two locators closed on other grounds. The **header** (`**Status:** GA` →
+`GA / Preview (Responsible AI scorecard)`) was applied by precedent, not re-asked — §9.13 settled
+the form. It needed no new citation: this file already states the preview fact in its own body at
+line 129, under a stamp already dated 2026-08-09. The parenthetical's *completeness* was verified
+rather than assumed — all seven `preview` occurrences grepped; six are the scorecard, and line 56
+is content inside a Transparency Note example, not a surface this file makes an availability claim
+about. The **dialect pair** (line 101 `model-` → `modell-`) aligned to the operator-ratified string
+at line 109.
+
+#### The replacement authored the defect, and was caught before commit
+
+The first draft of the Confidence line read `12 av 17 kilder verifisert via MCP (nr. 1–12), 5
+baseline (nr. 13–17)`. It would have **asserted that source 7 is MCP-verified** — and source 7 is
+itself labelled `(Status: Baseline — …)`. Writing it would have been the exact defect class the
+edit was closing, in the fix. The shipped wording counts block membership only
+(`17 kilder — 12 oppført under «Verified sources», 5 under «Baseline sources»`) and asserts nothing
+about any individual source, which leaves the finding below genuinely open instead of quietly
+asserted over.
+
+#### Booked rather than repaired
+
+- **idx-26s (new).** The two source blocks contradict themselves in exactly two places, in
+  *opposite* directions: #7 sits under **Verified sources** but is labelled `(Status: Baseline …)`;
+  #17 sits under **Baseline sources** but is labelled `(Status: Verified 2026-02 …)`. Not repaired,
+  because the direction of the fix is undecided and one direction renumbers the list. The shipped
+  count survives either way: by block membership 12/5, and swapping both mislabelled entries gives
+  verified {1–6, 8–12, 17} = 12 and baseline {7, 13, 14, 15, 16} = 5. **Eleventh instance of the
+  defect sitting next to the edit.**
+- **idx-26l … idx-26r (new, 7).** One entry per remaining affected file — 6 inconsistent plus the
+  1 ambiguous — each with a verbatim anchor verified unique before being written. Operator ratified
+  per-file entries over a single class-wide entry precisely because the queue schema carries one
+  `file` per entry: a class-wide entry would have left 20 files invisible to
+  `check-g7-queue.mjs`, which is the `795d494` failure repeated at class scale.
+- **`ai-red-team-operations-practical.md:87`** — the corpus's only other `model- og`. Left
+  untouched; the dialect fix was scoped to the one file.
+- **Untouched, on precedent:** `**Last updated:**` — now twelve ratified corpus edits, none of
+  which moved that field.
+
+#### No re-dating decision was required — and that was adjudicated before writing
+
+§9.13 established that re-dating a stamp is a scope decision taken up front, so the absence of one
+had to be *proved*, not assumed. Every stamp in the file sits at lines 59, 95, 131, 160, 312, 378 —
+all section-terminal, all before 378. Nothing after 378 carries a stamp except the inline
+`*(Verified MCP 2026-06-19)*` inside source 8, which is part of a source entry. Therefore: line 4
+sits above the first stamp, so the header edit falsifies none; the Kilder section and footer sit
+under no stamp at all; and line 101 sits under the section-3 stamp at 131, which is already dated
+2026-08-09 **and** claims verification against source rather than orthographic consistency — the
+entry says so itself. An orthographic alignment neither falsifies it nor requires re-dating it.
+
+**Queue:** 14 open / 13 resolved. **Suite:** 1047/1047.

skills/ms-ai-governance/references/responsible-ai/stakeholder-communication-ai-decisions.md

@@ -875,8 +875,6 @@ Hvis noen av disse mangler: **IKKE deploy før de er på plass.** AI uten stakeh
 
 **Total kilder**: 16 (9 verified fra MCP, 7 baseline/code samples)
 
-**MCP calls gjennomført**: 5 (3 docs_search, 2 docs_fetch, 1 code_sample_search)
-
 **Confidence vurdering**:
 - **Verified (90–100%)**: Azure ML Scorecard, Transparency Note, Interpretability, Responsible AI principles, AI governance structures
 - **High (75–90%)**: Generative AI explainability techniques, incident response patterns, EU AI Act framework

skills/ms-ai-governance/references/responsible-ai/transparency-documentation-standards.md

@@ -1,7 +1,7 @@
 # Transparency and Documentation - Regulatory and Best Practice Standards
 
 **Last updated:** 2026-06-19
-**Status:** GA
+**Status:** GA / Preview (Responsible AI scorecard)
 **Category:** Responsible AI & Governance
 **Type:** reference
 **Source:** https://learn.microsoft.com/azure/machine-learning/concept-responsible-ai
@@ -98,7 +98,7 @@ Strukturerte metadatabeskrivelser av AI-modeller og datasets. Originating fra ak
 
 ### 3. Responsible AI Scorecard
 
-PDF-rapport designet for å dele model- og data-innsikter mellom tekniske og ikke-tekniske stakeholders, spesielt for auditability og compliance.
+PDF-rapport designet for å dele modell- og data-innsikter mellom tekniske og ikke-tekniske stakeholders, spesielt for auditability og compliance.
 
 **Primære brukstilfeller:**
 
@@ -789,9 +789,8 @@ Return on investment: Transparency er billigere enn cleanup. Skal vi prioritere
     https://learn.microsoft.com/en-us/windows/ai/rai
     (Status: Verified 2026-02 — Model Cards reference, red teaming, governance processes)
 
-**Total MCP calls:** 5 (microsoft_docs_search: 3, microsoft_docs_fetch: 2, microsoft_code_sample_search: 1)
 **Unique sources:** 17 URLs
-**Confidence:** 80% Verified (MCP), 20% Baseline (established frameworks)
+**Confidence:** 17 kilder — 12 oppført under «Verified sources», 5 under «Baseline sources»
 
 ---