feat(ms-ai-architect): S1 judge bake-off RESULT — GATE FAIL (recall 68.4% < 80%) [skip-docs]
Blind fan-out: 15 batcher Opus 4.8 xhigh, 255 P-påstander dekket (0 missing/dupes/extra), hver judget mot live MS Learn via microsoft_docs_fetch (judgen så aldri gull-verdict). Resultat mot forhåndsregistrert gate (recall >=0.80, presisjon >=0.70): - judge: presisjon 81.3% (PASS), recall 68.4% (FAIL, Wilson [52.5-80.9%]), F1 0.743, 26/38 fanget - staleness baseline: recall 0/38 (judge slår den desisivt) - recall-drag konsentrert: sku 37.5%, taxonomy 66.7%; version/status/tpm allerede 80-86% - 10/12 bommer var grounded-men-feil (brittle dekomponering, isolert til sku/taxonomy) Per pre-registrert gate: STOPP, bygg IKKE S3. Fork (operatør) registrert i plan-doc: (a) honorer FAIL, (b) scoped judge, (c) målrettet iterasjon + re-kjør.
This commit is contained in:
parent
3e39f2df6b
commit
79e34312ea
4 changed files with 1666 additions and 0 deletions
1376
scripts/kb-eval/data/judge-bakeoff-results.json
Normal file
1376
scripts/kb-eval/data/judge-bakeoff-results.json
Normal file
File diff suppressed because it is too large
Load diff
Loading…
Add table
Add a link
Reference in a new issue