fix(ms-ai-architect): G7 idx-26g lukket — og §9.12s egen korpusmaaling var tatt over feil populasjon

Header **Status:** GA -> GA / Preview (Responsible AI scorecard). Begge kildesidene
bekreftet ordrett ved live fetch: 'currently in public preview ... provided without a
service-level agreement, and we don't recommend it for production workloads'. Begge
titlene sier '(preview)'. Formen valgt etter aa ha sjekket alle seks korpusfilene i
X / Preview (Y)-familien: parentesen navngir alltid det som ER i preview, aldri
GA-flaten.

DE TO BESLUTNINGENE VAR UAVHENGIGE, OG ENTRYEN BUNTET DEM. Banneret staar paa
concept-siden, som ALLEREDE er filas Kilder #1 - saa preview-faktumet trengte ingen
ny kilde. Bare de sju segmentene trengte how-to-siden. Delt opp i ratifiseringen slik
at en kildebeslutning ikke gated en status-retting.

STEMPELET RE-DATERT BEVISST 2026-08-04 -> 2026-08-09. En setning skrevet i dag under
et fem dager gammelt verifiseringsstempel er nettopp idx-26h-defekten ('stempelet var
FEIL DA DET BLE SATT, ikke foreldet') - denne gangen forfattet av editen selv.
Re-datering paastaar hele §1, saa hele §1 ble re-maalt mot begge live-sidene.

TREDJE LOCATOR, IKKE NAVNGITT AV ENTRYEN (ankere er en NEDRE grense, tiende gang):
Kilder #1 sa 'Status: GA (public preview for some features)' to linjer under sin egen
tittel som slutter paa '(preview)'. Banneret dekker hele funksjonen -> falskt, ikke
bare upresist.

Kilde #9 lagt til, 9-15 renummerert 10-16, Total kilder 15 -> 16. Trygt paa maaling:
grep over linje 1-794 gav NULL prosa-kryssreferanser til kildenummer. Formen bevarer
ogsaa telle-paastandens referent (teller entries, 16 entries / 9 URL-er), mens
'andre URL under #1' ville gitt 16 URL-er mot 15 entries - klassen idx-26j nettopp
bokfoerte.

RETTELSE: §9.12 (og idx-26j locator 5, og STATE) maalte **Status:** over ALLE 434
forekomster under skills/*/references/ - som blander header-feltet paa linje 4 med 47
SEKSJONSNIVAA-linjer i filkropper. To felt, to referenter, ett tall. Header-feltet
alene (FNR<=10): 387 forekomster / 389 filer, 69 distinkte. GA 249 (ikke 252),
Preview 1 (ikke 2), Komplett 1 (ikke 2). To-vokabular-funnet OVERLEVER rettelsen.
Konsekvensen gjoer det ikke: 'setter konvensjonen for 389 filer ved uhell' er FALSK -
69 distinkte verdier, ~44 singletons, 35+ sammensatte GA/Preview-former. En sann
sammensatt verdi SLUTTER SEG TIL etablert praksis. Paastanden som staar er smalere:
fila hadde alt valgt produkttilgjengelighets-aksen, og editen gjorde dens valgte akse
sann.

Bokfoert, ikke reparert: idx-26k (denne filas fottekst, 3+2+1=6 - samme klasse som
idx-26j locator 1, som BEKREFTER som maalt det 26j bare kunne kalle sannsynlig) +
idx-26j re-scopet (locator 5 er ikke lenger blokkert paa korpusbeslutning; formen er
naa presedens). **Last updated:** urort - verifisert at null av elleve tidligere
ratifiserte korpus-editer rorte feltet.

Koe: 8 aapne / 11 resolved. Suite 1047/1047.

[skip-docs]

docs/r11-pilot-results.md

@@ -1494,3 +1494,93 @@ file spans Transparency Notes (GA), Model Cards (practice) and the scorecard (pu
 preview) at once. **So idx-26g's header question is not local to either file — it is a
 corpus-wide field-semantics decision, and settling it one file at a time would set the
 convention by accident.**
+
+### 9.13 idx-26g closed — and §9.12's own corpus measurement was taken over the wrong population
+
+**The entry was right about the sources this time.** Both pages it names carry the Important
+banner verbatim — *"This feature is currently in public preview. This preview version is
+provided without a service-level agreement, and we don't recommend it for production
+workloads."* — and both titles read "(preview)". Fetched live 2026-08-09 rather than taken on
+the entry's word, because §9.12's lesson was that an entry can overstate its own evidence.
+
+**The two decisions were independent, and the entry bundled them.** The banner sits on
+`concept-responsible-ai-scorecard`, which is *already* this file's Kilder #1. So the preview
+fact needed **no new citation at all**; only the seven-segment enumeration needed
+`how-to-responsible-ai-scorecard`. Separating them meant the ratification question could put
+the a/b/c citation choice where it actually belonged instead of letting a citation decision
+gate a status correction.
+
+**Ratified:** header `**Status:** GA` → `**Status:** GA / Preview (Responsible AI scorecard)`,
+and the how-to page added as Verified source #9 with 9–15 renumbered to 10–16, `**Total
+kilder**` 15 → 16. The parenthetical form was chosen over the terser `varies by feature`
+after checking all six corpus files in the `X / Preview (Y)` family: in every one, the
+parenthetical names *what is in preview*, never the GA surface. Renumbering was safe on
+measurement, not assumption — grep over lines 1–794 found **zero** prose cross-references to
+source numbers. The chosen form also preserves the counting claim's pre-existing referent: it
+still counts numbered entries (16 entries, 9 carrying URLs), whereas the "second URL under
+Kilder #1" form would have left the file holding 16 URLs against 15 entries — the undefined-
+referent class idx-26j had booked one commit earlier.
+
+**The stamp was re-dated deliberately, because otherwise this edit would have authored the
+idx-26h defect.** §1's stamp read `2026-08-04`. Adding a sentence written on 2026-08-09
+underneath it would have produced text sitting under a verification date five days its senior
+— *"the stamp was false when it was set, not stale."* Re-dating claims all of §1, so all of §1
+was re-measured against both live pages: the seven segments map 1:1 to the how-to page's "How
+to read your scorecard"; `Du konfigurerer` matches the concept page's "target accuracy and
+target error rate"; the Formål bullets match its three unaddressed-needs bullets; the Verdi
+bullets match "building trust and gaining their approval for deployment"; and the YAML
+workflow block, the one piece that is synthesis rather than quotation, has every step grounded
+in "Who should use a Responsible AI scorecard?".
+
+**A third locator the entry did not name.** Kilder #1 read `Status: GA (public preview for
+some features)` — two lines below its own title ending in "(preview)". The banner covers the
+whole feature, so this was false rather than merely imprecise. Anchors are a lower bound,
+tenth instance.
+
+#### The correction: §9.12 measured `**Status:**` over the wrong population
+
+§9.12 (and idx-26j's fifth locator, and the STATE hand-off built on both) reported the field's
+value space as `GA` 252 · `Preview` 2 · `Komplett` 2 alongside `Established Practice` 43 ·
+`Gjeldende` 18 · `Reference` 7. Those counts were taken over **every** `**Status:**` occurrence
+under `skills/*/references/` — 434 of them — which silently merges the line-4 header field with
+**47 section-level `**Status:**` lines inside file bodies**. Two different fields, two different
+referents, one number.
+
+Re-measured on the header field alone (`FNR<=10`): **387 occurrences over 389 files, 69
+distinct values.** Corrected: `GA` **249** (not 252) · `Preview` **1** (not 2) · `Komplett`
+**1** (not 2). `Established Practice` 43, `Gjeldende` 18 and `Reference` 7 hold as stated.
+
+The two-vocabulary finding **survives the correction** — document maturity and product
+availability genuinely are mixed in one field, and that observation was right. What does not
+survive is the *consequence*: **"settling it one file at a time would set the convention by
+accident" is false.** The header field carries 69 distinct values over 387 files, roughly 44 of
+them appearing exactly once, and 35+ are compound GA/Preview forms. There is no single
+convention available to be set by accident; an accurate compound value *joins* an established
+practice. The claim that survives is narrower and does not need a corpus-wide decision to
+license the edit: **this file had already chosen the product-availability axis for its header,
+and the edit made its chosen axis truthful.** Choosing an axis for 389 files was never on the
+table.
+
+This is the §9.11 defect class turned on the *measurement* rather than the resolution — a
+premise measured over a population wider than the field it was reporting on. The cheapest
+ground-truth check (`FNR<=10`) was cheaper than the corpus-wide operator decision it would
+have triggered.
+
+#### Booked rather than repaired
+
+- **idx-26k (new).** This file's footer reads `**MCP calls gjennomført**: 5 (3 docs_search, 2
+  docs_fetch, 1 code_sample_search)` — components summing to 6. Same class as idx-26j's first
+  locator in the sibling, different wording and different field name. That **confirms as
+  measured** what idx-26j could only call likely: the footer form is replicated. Not repaired
+  here, because the referent is undefined in the same way (historical generation run, or
+  current provenance?) and idx-26j owns the form decision. Changing 5 to 6 would pick a
+  referent by stealth.
+- **idx-26j re-scoped.** Its fifth locator is no longer blocked on a corpus-wide decision: the
+  form is now precedent. Its header (`**Status:** GA`, spanning Transparency Notes GA, Model
+  Cards practice, scorecard preview) is false the same way, and a divergence from the ratified
+  form would now be *introduced* by that entry rather than inherited.
+- **Untouched, on precedent:** `**Last updated:**` — verified that zero of the eleven prior
+  ratified corpus edits modified that field. `Verifisert: 2026-02` on sources 1–8 — those pages
+  were not re-fetched.
+
+**Queue:** 8 open / 11 resolved. **Suite:** 1047/1047.

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

@@ -1,7 +1,7 @@
 # Stakeholder Communication - Explaining AI Decisions to Non-Technical Audiences
 
 **Last updated:** 2026-06-29
-**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
@@ -49,6 +49,8 @@ Hver gruppe krever tilpasset kommunikasjon: ledere trenger risikovurdering, slut
 
 **Primært bruksområde**: Azure Machine Learning
 
+**Tilgjengelighet**: Scorecard-en er i public preview. Preview-versjonen leveres uten SLA, og Microsoft anbefaler den ikke for production workloads. Enkelte funksjoner kan mangle støtte eller ha begrenset funksjonalitet.
+
 **Formål**:
 - Bygge bro mellom tekniske verktøy og etiske/regulatoriske krav
 - Muliggjøre multi-stakeholder alignment i ML-livssyklusen
@@ -82,7 +84,7 @@ Business Stakeholder: Vurderer om modellen møter forretningskrav → Go/No-go b
 - ✅ Dokumentasjon som kan deles med juridisk og compliance
 - ✅ Grunnlag for deployment-godkjenning
 
-*Confidence: Verified (MCP microsoft-learn, 2026-08-04)*
+*Confidence: Verified (MCP microsoft-learn, 2026-08-09)*
 
 ---
 
@@ -798,7 +800,7 @@ Hvis noen av disse mangler: **IKKE deploy før de er på plass.** AI uten stakeh
 
 1. **Share Responsible AI insights using the Responsible AI scorecard (preview)**
    - https://learn.microsoft.com/en-us/azure/machine-learning/concept-responsible-ai-scorecard?view=azureml-api-2
-   - Status: GA (public preview for some features)
+   - Status: Public preview — leveres uten SLA, ikke anbefalt for production workloads
    - Verifisert: 2026-02
 
 2. **What is Responsible AI?**
@@ -836,37 +838,42 @@ Hvis noen av disse mangler: **IKKE deploy før de er på plass.** AI uten stakeh
    - Status: GA
    - Verifisert: 2026-02
 
+9. **Use Responsible AI scorecard (preview) in Azure Machine Learning**
+   - https://learn.microsoft.com/azure/machine-learning/how-to-responsible-ai-scorecard
+   - Status: Public preview — leveres uten SLA, ikke anbefalt for production workloads
+   - Verifisert: 2026-08-09
+
 ### Baseline Sources (modellkunnskap)
 
-9. **EU AI Act** (via EØS-avtalen, relevant for Norge)
-   - Confidence: High (publicly available regulation)
+10. **EU AI Act** (via EØS-avtalen, relevant for Norge)
+    - Confidence: High (publicly available regulation)
 
-10. **NIST AI Risk Management Framework**
+11. **NIST AI Risk Management Framework**
     - Confidence: High (US standard, widely referenced)
 
-11. **Norsk offentlig sektor AI governance** (DFØ, Datatilsynet)
+12. **Norsk offentlig sektor AI governance** (DFØ, Datatilsynet)
     - Confidence: Medium (basert på generell kunnskap om norske myndigheters krav)
 
-12. **Power Platform CoE Starter Kit**
+13. **Power Platform CoE Starter Kit**
     - Confidence: High (open-source, dokumentert av Microsoft)
 
 ### Code Samples (fra MCP microsoft_code_sample_search)
 
-13. **MLflow GenAI evaluation scorers**
+14. **MLflow GenAI evaluation scorers**
     - Eksempler på å evaluere AI-responder med custom judges
     - Relevant for: Quality assessment og stakeholder-rapportering
 
-14. **Azure AI tracing with OpenTelemetry**
+15. **Azure AI tracing with OpenTelemetry**
     - Eksempler på å logge AI interactions med feedback
     - Relevant for: Audit trail og user feedback loops
 
-15. **Azure AI Evaluation SDK**
+16. **Azure AI Evaluation SDK**
     - Eksempler på å bruke built-in evaluators (RelevanceEvaluator, ViolenceEvaluator)
     - Relevant for: Safety og quality metrics for stakeholders
 
 ---
 
-**Total kilder**: 15 (8 verified fra MCP, 7 baseline/code samples)
+**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)