feat(ms-ai-architect): ratifiser form #35 (motsier-kilden), trekk idx-33i fra #34, og foer populasjonsregel-beslutningen [skip-docs]

Tre operatorbeslutninger tatt 2026-08-25, etter adjudiseringen i 02b9c2a.

1) FORM #35 RATIFISERT - motsier-kilden-klassen. Fire medlemmer: idx-33k
(grunnleggeren, reist 20.08 og bevisst holdt utenfor #34) + idx-33q/r/s, som
er EN defekt i tre filer fordi tre filer siterer samme side.

idx-33k ble RE-VERIFISERT mot fersk henting for formen ble skrevet - nettopp
fordi idx-33i viste at et medlem maalt en gang og latt staa apent kan vaere en
falsk positiv. Kilde per passasje ble ogsaa re-etablert paa seksjonsniva
foerst. Levende action-groups 2026-08-25: "Azure Function | Managed Identity
Support: Yes". Fila sier No paa L76 og navngir Azure Function blant de som
ikke stotter det paa L84. Fire typer stotter MI, fila navngir tre. Funnet
staar. Samme fil hviler forovrig paa en ANNEN kilde i en annen seksjon (L41 ->
alerts-processing-rules, adjudisert kildetro i dag).

#35 baerer to regler #34 ikke har: reparer den gale klausulen og ENDRE
INGENTING ANNET (i alle tre window-size-tilfellene ligger defekten som ett
tall i en ellers ordrett tro passasje), og MAAL KILDEN KORPUS-BREDT FOER
ANVENDELSE - #29 anvendt per KILDE i stedet for per klasse.

2) #34 RETTET - idx-33i trukket fra GOVERNS og fra gren A. Medlemstallet gikk
fra atte til sju, og de tre setningene som baar det tallet er rettet med.
Grammatikk-argumentet staar naa paa idx-33h alene, og tilbaketrekkingen er
skrevet inn som en STYRKING av regelen: et "Kun" sa ingenting om grenen, og
det viste seg aa si ingenting om hvorvidt det i det hele tatt var en defekt.
Tilbaketrekkings-notatet ligger paa EGEN linje, aldri paa GOVERNS-linja, saa
en grep paa den linja aldri kan returnere et medlem formen ikke styrer.

3) POPULASJONSREGELEN BEHOLDES - fort inn i nett-artefaktet saa ingen "fikser"
den senere som en apenbar bug. Lekkasjen (naermeste overskrift er en
UNDERoverskrift) slapp inn 5 av de 29; fire ble kildetro, den femte er
idx-33o, et ekte medlem. Aa stramme regelen har altsaa en MAALT KOSTNAD.
Samme artefakt: sample_is_not_random sa i presens at 21+13 kandidater er
umaalte - det ble usant i dag og er merket SUPERSEDED, og nettet er ført som
UTTOMT.

Den ventende ordren 20260825T070602Z-862318346 er oppdatert i samme slengen:
den ble skrevet FOER disse beslutningene og baar to premisser som flyttet seg.
Instruksen (ikke ror dem) er uendret; begrunnelsene er rettet.

Ingen KB-fil rort. check-g7-queue 0 drift, suite 1052/1052.
Diff: 6/5 i koa, 4/2 i nettet - ingen reformatering.

scripts/kb-eval/data/idx-33-net-2026-08-20.json

@@ -4,8 +4,10 @@
     "why_tracked": "The 2026-08-13 and 2026-08-18 nets both lived in a scratchpad and were both gone by the next session, so the class had to be re-measured from scratch twice. This file is the third attempt made durable. It is DATA, not a gate: nothing validates it and no test reads it.",
     "not_reproducible_from_18_08": "Same corpus (283 stamps / 89 files) but this run's subtraction rules and vocabulary differ from 2026-08-18: 42 source-section + 18 link-entry subtractions -> 223 claim passages (18.08 measured 37 + 20 -> 226), and 32 container / 19 limiter (18.08: 16 / 26). Treat 18.08's numbers as unreproducible, not wrong.",
     "known_positive_control": "Run BEFORE any flag was read. ai-threat-modeling-stride.md:210 -> container (\"across Azure subscriptions\"); digital-accessibility-action-plan.md:160 -> container (\"for K-12 laerere\"). Both PASS. A net that fails this control proves nothing.",
-    "sample_is_not_random": "The 17 stage-2 adjudications were hand-picked toward enumerations and explicit limiters - the stratum where narrowing is even possible. They carry NO rate. 21 of 32 container and 13 of 19 limiter candidates were never taken to a live source and are UNMEASURED, not clean.",
+    "sample_is_not_random": "The 17 stage-2 adjudications were hand-picked toward enumerations and explicit limiters - the stratum where narrowing is even possible. They carry NO rate. 21 of 32 container and 13 of 19 limiter candidates were never taken to a live source and are UNMEASURED, not clean. SUPERSEDED 2026-08-25: that sentence was true when written and is no longer. All 29 genuinely unmeasured candidates (34 without a stage-2 verdict, minus 5 that already had one from the 2026-08-18 round) were taken to live microsoft-learn sources on 2026-08-25 under order 20260825T063639Z-3599226074. See exhausted_2026_08_25 below.",
-    "population_rule": "For each line containing \"Verified MCP\": the passage is that line plus following lines until a blank line, a heading, or the next stamp. A passage is excluded if its nearest preceding heading matches /kilde|source|referanse|reference|further reading|videre lesning/i, or if the stamped line is itself a markdown link list entry."
+    "population_rule": "For each line containing \"Verified MCP\": the passage is that line plus following lines until a blank line, a heading, or the next stamp. A passage is excluded if its nearest preceding heading matches /kilde|source|referanse|reference|further reading|videre lesning/i, or if the stamped line is itself a markdown link list entry.",
+    "exhausted_2026_08_25": "THIS NET IS EXHAUSTED. Every one of the 51 candidates now carries a verdict: 17 adjudicated 2026-08-20 (recorded in stage2 and in idx-33c), 5 adjudicated 2026-08-18, and the remaining 29 adjudicated 2026-08-25 (commit 02b9c2a). Outcomes over those 29: 19 faithful to source, 5 members of the under-claim class governed by form #34 (idx-33l, idx-33m, idx-33n, idx-33o, idx-33p), 4 of another class (idx-33q/idx-33r/idx-33s, one defect in three files, plus idx-33t), 0 not adjudicable, and 1 that is not a claim passage at all. The full batch record - denominator, outcome split, false-positive count and the source-per-passage findings - lives ONCE, in the summary of queue entry idx-33l, and is deliberately not restated here. Do not re-run this net expecting new candidates; a NEW net would need a different vocabulary or a different population rule, not another pass with this one.",
+    "population_rule_decision_2026_08_25": "OPERATOR DECISION 2026-08-25: THE POPULATION RULE STAYS AS WRITTEN. It is a measured leak, not an oversight, and it is recorded here so nobody \"fixes\" it later as an obvious bug. The rule subtracts a passage whose NEAREST preceding heading matches the source-section pattern. When a source section is split by a sub-heading - \"## Kilder og verifisering\" followed by \"### Microsoft Learn (Verified via MCP)\" - the nearest heading is the sub-heading, the pattern does not match, and the passage stays in the population. Five of the 29 candidates adjudicated on 2026-08-25 entered this way: endpoint-health-and-capacity-planning 629 and 633, real-time-streaming-monitoring 547, security-and-audit-logging-ai 408, and data-residency-audit-monitoring 558. FOUR adjudicated faithful-to-source; the fifth became idx-33o, a genuine member of the #34 class. So tightening the rule to test every heading upward has a MEASURED COST as well as a benefit: it would have excluded a real finding, and the 2026-08-25 batch also showed that source-list annotations make repairable claims (idx-33u was raised against one, and the 2026-08-25 session before it repaired another at agent-365 373). The rule is therefore kept deliberately, with the leak understood, rather than tightened on the assumption that a source-section passage cannot carry a defect."
   },
   "counts": {
     "stamps": 283,