Summary

The Option A wave (user-approved): kills the last blocking canary red and the answer-side half of the run-#60 monitoring targeting misses. Two single-concern changes plus a review-response hardening commit:

0abf3c9 — Option A: title-supported escalation rescue (protected retrieval surface). The neuroleptic-side-effect-escalation case failed the citation gate on every canary (#57→#60): the query misclassifies as medication_dose_risk, its honest lexical pool fails the dose fast-path floor (live-measured 0.184/0.0125 vs 0.66/0.055), the S3 document-lookup title layer was excluded by the class allowlist, and the vector leg injects the wrong sibling doc (Zuclopenthixol). shouldAttemptDocumentLookupFastPath now also engages the layer for medication_dose_risk when BOTH existing deterministic classifier signals hold: intent === "escalation_risk" (assigned only when drug_dosing wording did NOT match — pure dose/route/frequency questions structurally cannot fire) and documentTitleTerms > 0 (a curated title alias phrase is present for the alias tier to rescue with). Rescue semantics by construction: the S3 block sits after both fast-path returns, so a firing query has already had its pool floor-rejected. Blast radius proven twice over: exactly one eval case fires — my executed sweep and the retrieval reviewer's independent 110-case classifier probe agree. Honest live-expansion disclosure: real shapes like "When should clozapine be stopped?" now engage S3 (surfacing the title-named curated SOP for the named entity — assessed clinically correct; the principal thing the dedicated canary pair watches). No imputation formula, released-comparator key, selection clamp, or threshold is touched.

1aebf02 + a3b9a54 — monitoring evidence gate parity (answer-side). Run-#60's four monitoring targeting misses traced to the answer-intent evidence gate being narrower than the fact filter and the eval judge: inflected schedules ("reviewed annually", "monitored for 3 hours"), bare digit durations, and level-range/metabolic-panel vocabulary passed fact filtering but died at the gate, at kind classification, or at the intent-token coverage checks. One shared monitoringScheduleEvidencePattern now feeds both the gate and the filter (lockstep by construction), and both coverage checks gain a figure escape mirroring the concrete-dose escape, scoped to text carrying the asked-for interval or unit range itself. The review-response commit a3b9a54 closes both reviewer-found hazards: the inflected kind tokens no longer steal sole-dose-value sentences from the dose arm (legacy tokens classify byte-identically; new-inflection-only sentences fall through to dose when they carry a concrete dose value — both reviewer repro sentences pinned as tests), and the multi-drug bare-row guard extends to monitoring_schedule so a schedule row naming no drug inside a multi-drug chunk cannot inherit the queried drug (both governance-traced misattribution paths now refused, red-proven). Expected live effect: quality-lithium-monitoring-range + quality-antipsychotic-metabolic-monitoring gain their figures (monitoring targeting ≥3/5); the other two misses are retrieval-depth (documented residual).

All four reviewer subagent verdicts (pre-ready, full ledger rows on branch):

• rag-retrieval-reviewer on parity: REQUEST-CHANGES (soft) — reproduced the dose-steal P2 (fails closed) + multi-drug P3 → both fixed in a3b9a54, red-proven both directions.
• clinical-governance-reviewer on parity: APPROVE-WITH-NITS — its independently-converged P2 (multi-drug cross-entity attribution) is the same hazard, fixed in a3b9a54 exactly as its smallest-fix prescribed; unsupported-figures/conservative-failure/adjacent-context/PHI all verified safe.
• rag-retrieval-reviewer on rescue: APPROVE-WITH-NITS, no P0/P1/P2 — independent 110-case probe confirms exactly one firing case; binding constraints byte-identical; downstream inert (empty requiredTermSignals; wrong-med cap never engages); S3 fixture shapes verified faithful to production values; redundant-optional-chain nit taken (55a6bc3).
• clinical-governance-reviewer on rescue: APPROVE-WITH-NITS, no P0/P1 — wrong-document impossible (alias phrases must appear in the query; per-document groups), purely additive availability, adversarial fail-closed double layer executed, governance metadata unbypassed; P2 = the live canary trust gate below.

RAG impact: behaviour change — canary pair: run #60 (banked baseline on main 9b655fa) -> dedicated post-merge dispatch (answer_case_limit=44 + answer_quality_eval=true, ~$2-4, authorized envelope); success = citation_failure_rate 0.0227 -> 0, golden doc/content recall pinned 1.0/1.0, zero per-case rr@10 regressions, monitoring targeting ≥3/5, no other medication_dose_risk case's top-5 shifting; regression = single-commit revert of the offending commit per the proven drill.

Verification

• Red-proofs both directions: parity — four tests verified failing pre-change; review-response — dose-steal repros + multi-drug guard verified red with fixes reverted; rescue — predicate + end-to-end ordering tests verified failing with the gate reverted; negative guards pin pure-dose/no-title/non-escalation shapes and the four allowlisted classes
• Focused: escalation-rescue 8/8; imputation contract + fast-path-ordering + retrieval-selection + second-stage + ranking-tuning + eval-retrieval 143/143; extractive suites 147/147; reviewer-run batches 644/644 + 31/31
• npm run verify:cheap — all static checks green; unit suite 3087 passed / known container-only pdf artifact
• typecheck, prettier, maintainability budgets (rag.ts 4919/5030)
• check:production-readiness — fails only on absent secrets in this demo container (expected per repo docs); hosted CI runs the CI-safe variant
• Dedicated live canary pair post-merge (criteria in the RAG impact line)

Risk and rollout

• Risk: medium — deliberate retrieval behaviour change on a protected surface, bounded by predicate narrowness (two ANDed deterministic classifier signals; two independent sweeps show one firing eval case), rescue-only semantics (S3 candidates ADD to the pool; previously-returned results remain retrievable — test-pinned), and the unchanged floor/selection/release contracts. The answer-side widen is bounded by the unchanged promotion guard + claim-support + numeric gates plus the new dose-steal and multi-drug guards (worst case stays a conservative evidence gap, never a wrong or misattributed figure).
• Rollback: single revert per commit; no schema/env/RPC/cache changes.
• Provider or production effects: the post-merge canary pair dispatch (~$2–4, user-authorized envelope).

Clinical Governance Preflight

• Source-backed claims still require linked source verification before clinical use — both changes sit upstream of the unchanged claim-support/numeric/citation gates
• No patient-identifiable document workflow was introduced or expanded
• Supabase target remains Clinical KB Database (sjrfecxgysukkwxsowpy) — no access changes
• Service-role keys and private document access remain server-only
• Demo/synthetic content remains clearly separated from real clinical sources
• Source metadata, review status, and outdated/unknown-source behavior remain conservative — S3-surfaced documents flow through the same governance annotation as all results
• Deployment classification/TGA SaMD impact was checked: escalation queries gain the title-named curated SOP as top source; monitoring answers gain verbatim-supported, correctly-attributed schedules — both directions are toward the clinically correct document/figure, live-pair-validated before trust

🤖 Generated with Claude Code

https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9

Summary by CodeRabbit

• New Features
  • Improved monitoring_schedule evidence understanding with inflection-tolerant phrasing (cadence/follow-up/baseline) and direct interval/range figure coverage.
  • Added a conditional “escalation rescue” for medication_dose_risk when escalation intent and matching document title terms are present.
• Bug Fixes
  • Prevented monitoring evidence from being derived from dose-bearing sentences; expanded safe “bare row” handling for monitoring_schedule.
  • Strengthened monitoring-gate parity to admit valid interval/range facts without breaking stable, order-consistent retrieval.
• Tests
  • Added monitoring evidence gate parity and escalation rescue/order-stability regression coverage; refined figure-coverage expectations.
• Documentation
  • Updated the branch review ledger with new review outcomes and gates.