feat(rag): figure-aware extractive selection for dose/monitoring answers (E-3c PR-C)

docs/branch-review-ledger.md

@@ -670,3 +670,5 @@ Use this ledger to prevent repeated branch and PR reviews when the reviewed HEAD
 | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (PR: E-3b budget-aware generation deadlines) | 314d03f (+eb08ec5 review row) | ADDENDUM 5 wave E-3b implemented per the design-agent plan: generation attempts clamped to route budget minus a measured 2s recovery reserve (single call-site, all four attempt kinds); truncation self-heal gated on retry viability (reserve+5s floor) with observable truncation_retry_skipped_budget_reserve marker; additive route_budget_exhausted_by_retrieval runtime flag; eval route-ceiling gains the triple-condition cross-region carve-out (context + runtime flag + zero generation) with retrieval-exhausted audit cells — local/release gates provably strict. Fixes I3 (54ms budget overrun after 22.6s provider timeout), I5 (82s truncation waste class), resolves I2 (clozapine 13.3s retrieval vs 12s runtime budget = geography, now suppressed ONLY in the sanctioned cross-region context with full auditability). Reviewer verdicts: rag-retrieval-reviewer APPROVE-WITH-NITS (2 P3: prod-dead requestTimeoutMs retained for symmetry; report-cell coupling cosmetic; cached-replay invariant PROVEN — budget-exhausted answers never cached, carve-out unreachable via replay; marker isolation proven — SLO counters key on fallback_reason not answer_retry_reasons); clinical-governance-reviewer APPROVE-WITH-NITS (prior row) — internal-timeout→PublicApiError→existing-fallback path verified, all safety gates still applied to recovery answers. ALSO BANKED — E-2 targeting baseline (canary run #58, 29788404357, all-green incl. first execution of the !cancelled()-fixed instrument, ~$1-2): metric_rates relevance 0.6 / readability 1.0 / artifact_leaks 1.0 / intent_coverage 0.9333 / fail_closed 0.9; targeting_rate 0.5909 (13/22); by intent: document_lookup 5/5, red_result_action 3/3, contraindication 2/2, dose 1/5, monitoring_schedule 1/5, pathway_referral 1/2; all 9 misses = missing dose figure/schedule-interval (answer lengths 73-232 chars) → E-3c co-primary target alongside the wasted-generation class. Phase E spend ≈$3-6 of ≤$20. | Red-proofs: reserve pinned 3 independent ways (exact 23000ms grant, deadline flag clear, total under budget); self-heal skip pins exact marker + single provider call; offline flag pinned true/false. Focused: route-budget 9/9, eval-quality 27/27, fallback+offline 52/52, parser/abort regressions 15/15. Full suite 3043 passed / 1 known container pdf artifact. typecheck+lint+prettier clean. No provider calls; live proof = E-4 paired run |
 | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (restarted from e02ba3d; PR: E-3c PR-A instrument) | see PR head | ADDENDUM 5 wave E-3c design accepted (Plan agent, full report in session record) and PR-A delivered: eval-answer-quality gains --dump-answers <path> (per-case answer TEXT + sections + targeting verdict JSON for the canary artifact — answers are not retained at rest by privacy design, so miss diagnosis needs eval-time capture), parseArgs/buildAnswerDumpRecord exported behind an import.meta main-guard (eval-quality precedent), canary targeting step wired with the flag. Design highlights for the record: DEFECT 1 root confirmed as fast-attempt-doomed-then-discarded on strong_routine_retrieval procedural shapes — PR-B generalizes the EXISTING validated-extractive short-circuit pattern (LAI + blocked-recovery precedents, hasValidatedExtractiveCandidate) to the measured shape via a new rag-extractive-first.ts module (net ~-110 rag.ts lines against the 5030 budget); DEFECT 2 rank-1 root = extractive lead-slot selection prefers shortest sentence and admits figure-less leads (sort at rag-extractive-answer.ts:917, 1-slot monitoring leads) — PR-C adds intent-figure-aware lead promotion with a claim-support atom-corpus nuke-proofing guard + dose/threshold fallback candidate preference (find(safe && figure) ?? find(safe)); answer-verification CLEARED as direct cause (whole-answer gap or unbold only, never per-figure deletion). H2 (strong-route comparison/complex residual) explicitly deferred as the named E-3d candidate. E-4 metrics set: discarded-generation <20% from ~47%, dose ≥3/5, monitoring ≥3/5, no intent below #58, recalls pinned 1.0, relevance/fail_closed/readability/artifact_leaks ≥ #58. | PR-A gates: new tests/eval-answer-quality.test.ts 4/4; typecheck, check:github-actions, check:ci-scope, eval-canary-workflow test, prettier all clean; no provider calls |
 | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (PR: E-3c PR-B short-circuit) | f7e6cbb + hardening commit | E-3c PR-B: pre-generation validated-extractive short-circuit for gate-passed routine procedural "What...process/include/required" queries (marker validated_routine_extractive_first), generalizing the LAI + blocked-recovery precedents via new rag-extractive-first.ts (3 predicates moved byte-verbatim, machine-verified; rag.ts 5029→4908 vs 5030 budget). Kills the run-#57 6x wasted-generation class. REVIEWS (both pre-push): rag-retrieval-reviewer APPROVE-WITH-NITS — move fidelity brace-diff verified byte-identical; confidence-gate skip PROVEN safe (passed-gate is a no-op in applyConfidenceGate; markers pairwise mutually exclusive); comparison false-positives blocked by unchanged classifier precedence; offline/source-only idempotent; zero retrieval/ranking/selection/threshold change; P2 = eval-only assertions (intent_coverage/artifact_leaks/expected-file) unverifiable offline for flip candidates (quality-nocc-document-support, quality-form-required-documentation, quality-discharge-documentation, quality-duress-pathway + rag-set siblings) → pre-merge BRANCH canary recommended and ADOPTED (offline-green + review-approved proven insufficient for this surface, 2026-07-20). clinical-governance-reviewer APPROVE-WITH-NITS — full gate-stack trace: nothing bypassed (same finalizeRagAnswerQuality, same citation scoping, numeric verification not fail-open, ungrounded-finalize defense at rag.ts:3694); P2 = pre-existing bare-cross-reference-with-overlap gap, NOT materially widened (new trigger anti-correlates), hardening recommended → APPLIED this PR: !isBareCrossReferenceAnswer screen in hasValidatedExtractiveCandidate (closes all three short-circuit paths; discriminating test added; disclosed post-review delta, strictly narrows shipping). MERGE GATE: draft until the branch canary pair (baseline #57/#58 vs branch run with answer_case_limit=44 + answer_quality_eval=true, est $3-6 of authorized envelope) is green — zero per-case regressions, recalls 1.0, quality/targeting rates >= baseline. | Red-proof + 3 negative guards; focused 61/61 + fallback/offline/contract suites; full suite 3061 passed / 1 known container artifact (pre-hardening tree; hardening re-verified focused); typecheck+lint+prettier clean; maintainability budget passed (4908/5030) |
+| 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (pair verdict — no code change; main stays at 22b6a2e) | canary run 29794759627 (#59, branch head 7310cb3 = merged PR-B content) | E-3b + E-3c PR-B LIVE-VALIDATED (pair vs banked #57/#58): route_ceiling_failures 2→0 (E-3b proven: agitation timeout now fits inside budget; clozapine retrieval-exhausted ceiling honestly excused via the triple-condition cross-region carve-out — exactly one "retrieval-exhausted" audit cell in the report); p95 17.4s→15.47s (-11%); golden retrieval SUCCESS 36/36 (stop-ship criterion held); validated_routine_extractive_first fired on 4 of the 6 target cases (patient-safety-plan 3.3s, treatment-team-process 2.6s, ect-procedure 2.2s, illegal-substances 2.2s — all pure extractive, zero generation, was 6-9s each with a discarded attempt), the other 2 (community-home-visits, best-practice-prescribing) stayed on generation+fallback because their extractive candidates legitimately fail validation gates = the designed-conservative outcome; ZERO new failing cases (list 5→2, both known residuals: neuroleptic citation red = Option A territory, admission-comparison expected-doc = non-blocking labeling residual); grounded 1.0 + unsupported_correct 1.0 held; targeting 0.5909→0.619, fail_closed 0.9→0.9333, readability/artifact_leaks/intent_coverage unchanged; relevance 0.6→0.5667 = single-case wobble on n=30, WATCH in E-4, not a gate. Discarded-generation rate materially down (4 conversions; residual = the H2 strong-route slice named as E-3d candidate, per design's 20-33% expectation band). MERGE STANDS (user had armed auto-merge pre-verdict; revert drill not triggered). Spend +~$3-6 → Phase E total ~$6-12 of ≤$20. | Pair evidence: run #59 job log (Blocking failures = citation only; Answer Case Diagnostics markers; metric_rates + targeting blocks); dump artifact populated for PR-C diagnosis |
+| 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (PR: E-3c PR-C figure-aware selection) | 043b030 + P2-fix commit | E-3c PR-C: dose/monitoring extractive answers now carry the asked-for figure/schedule when the cited chunk verbatim supports it — lead-slot promotion (dose swaps last of 2 slots, monitoring appends 2nd sentence; no-op when a lead already carries a figure) guarded by the claim-support atom corpus (sourceEvidenceText exported, promotionAtomKey byte-identical to claim-support's atomKey), plus the dose/threshold generation-fallback preferring the safe figure-carrying candidate (safety gate unchanged, filter-order-stable). Fallback helpers extracted to rag-extractive-answer (cycle-check verified); rag.ts 4908→4901. Six discriminating tests each verified red-on-prior-code incl. proving the nuke-guard load-bearing by disabling it. REVIEWS (both pre-push on 043b030): rag-retrieval-reviewer APPROVE-WITH-NITS — no-op path byte-identical verified, atom-key identity verified, filter-vs-find proven side-effect-free, 2-sentence append gate-safe, intent double-gated, zero retrieval/ordering change, imputation contract green; P2 = zero-atom monitoring figures ("every 6 weeks" yields no value atom) pass the guard trivially and can be nuked by claim support if sourced only from adjacent context (fails SAFE — evidence gap, never a wrong figure). clinical-governance-reviewer APPROVE-WITH-NITS — all six clinical concerns CLEARED end-to-end (verbatim-support guarantee, citation binding preserved, conservative failure test-proven, unsafe candidates impossible, wrong-drug risk controlled by pre-existing entity/multi-drug guards, no PHI); same zero-atom finding as P3 + one comment-precision nit. P2 FIXED post-review (disclosed): zero-atom figures now require the matched figure substring verbatim in sourceEvidenceText (intentFigureMatchText); proven both directions by 3 new tests (promotes from content, refuses from adjacent-context-only); comment-precision nit folded in. | Focused post-fix: extractive-formatting 32/32 + fallback/eval-cases/offline/contract/extractive-first 98/98 incl. imputation contract; typecheck+prettier+budgets clean; full-suite 3068-passed baseline pre-P2-fix (fix re-verified focused). Live proof = E-4 pair next |

src/lib/rag/rag-claim-support.ts

@@ -99,7 +99,12 @@ function splitComparisonClaims(value: string) {
     .slice(0, 24);
 }
 
-function atomKey(atom: ClinicalValueAtom) {
+/**
+ * Canonical identity for a clinical value atom, shared with the extractive
+ * lead-figure promotion guard (rag-extractive-answer.ts) so promotion and
+ * claim-support agree on whether a figure is present in the evidence corpus.
+ */
+export function clinicalValueAtomKey(atom: ClinicalValueAtom) {
   return [
     atom.kind,
     atom.canonicalValue,
@@ -212,7 +217,14 @@ function compatibleSafetyDimensions(claim: string, evidence: string) {
   return [...safetyDimensions(claim)].every((dimension) => evidenceDimensions.has(dimension));
 }
 
-function sourceEvidenceText(source: SearchResult) {
+// Exported so the extractive figure-promotion guard (rag-extractive-answer.ts) can check a
+// candidate fact's value atoms against the EXACT corpus this module assesses claims with.
+// Keep the two in lockstep: a promoted figure verified against a wider corpus (e.g. one that
+// includes adjacent_context) would pass numeric verification and then trip
+// claim_support_high_risk_gap here. (Precision note: this corpus also includes
+// index_unit metadata that the numeric-verification corpus does not, so it is not a strict
+// subset — that sole divergence fails safe: the guard passes, the numeric gate then nukes.)
+export function sourceEvidenceText(source: SearchResult) {
   return [
     source.section_heading,
     source.content,
@@ -243,8 +255,8 @@ function sourceSupportsClaim(claim: string, source: SearchResult) {
   if (!compatibleSafetyDimensions(claim, evidence)) return false;
   if (!compatibleHighRiskTrigger(claim, evidence)) return false;
 
-  const evidenceAtoms = new Set(extractClinicalValueAtoms(evidence).map(atomKey));
-  if (extractClinicalValueAtoms(claim).some((atom) => !evidenceAtoms.has(atomKey(atom)))) return false;
+  const evidenceAtoms = new Set(extractClinicalValueAtoms(evidence).map(clinicalValueAtomKey));
+  if (extractClinicalValueAtoms(claim).some((atom) => !evidenceAtoms.has(clinicalValueAtomKey(atom)))) return false;
 
   const claimTopics = topicTokens(claim);
   const evidenceTopics = topicTokens(evidence);