2462685ab8
Some checks failed
CI / lint (push) Successful in 1m32s
CI / notebooks-smoke (push) Successful in 2m46s
Deploy / changes (push) Successful in 35s
CI / test (push) Failing after 8m25s
Infra CI / notebooks (push) Successful in 3m22s
Infra CI / zotero (push) Successful in 51s
Infra CI / docs (push) Successful in 49s
Infra CI / api (push) Successful in 5m16s
Infra CI / bcda (push) Successful in 5m44s
Infra CI / mc (push) Successful in 57s
Deploy / notebooks (push) Has been skipped
Deploy / zotero (push) Has been skipped
Deploy / docs (push) Has been skipped
Infra CI / llm (push) Successful in 2m49s
Deploy / api (push) Successful in 2m12s
Deploy / bcda (push) Successful in 2m30s
Deploy / mc (push) Has been skipped
Deploy / llm (push) Successful in 1m51s
Deploy / report (push) Successful in 41s
Vocab v3 adds six hospital-side themes the golden set showed missing: opps-asc-payment, ipps-drg-payment, hospital-wage-index, non-opioid-pain (NOPAIN), hospital-quality-programs and hospital-conditions-of-participation. Every v2 theme's text is unchanged and the file declares compatible_with: [2]; the tagger reuses all stored v2 verdicts and judges a new theme only when it enters a shortlist, so the corpus need not be re-judged. Golden v3 is golden v2 relabelled for v3 (50 entries; 3 genuinely off-topic comments remain). Live, 14b, top 12, floor 0.45: v2 held-out F1 0.74 (P 0.95 R 0.61, abstain 23%) v3 held-out F1 0.76 (P 0.89 R 0.67, abstain 14%) - gate PASS IPPS 0.29 -> 0.70, OPPS 0.27 -> 0.70; tune F1 0.81. Two new definitions were narrowed after both halves showed the same false positives (CoP on hospital-quality comments, NOPAIN on opioid-sparing letters). stack llm eval-tags with no flags now runs that gate: golden v3, held-out half, micro-F1 >= 0.75, per-theme precision >= 0.6; a golden file without halves is scored whole.