feat(ci): weekly canary guards the full 44-case answer set (I9)#1044
Merged
Conversation
…se canary Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9
The scheduled Sunday canary's ANSWER_CASE_LIMIT default rises 8→44 now that run #61 proved the citation gate green on the full set — both historical blocking reds lived outside the first 8 cases, leaving the weekly gate blind to them. Dispatch input override unchanged; contract test pin updated in lockstep. Est +$1-2/week, plan-authorized (ADDENDUM 5 I9, post-green). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9
Contributor
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (3)
📝 WalkthroughWalkthroughThe eval-canary workflow’s default and fallback ChangesCanary case limit
Estimated code review effort: 1 (Trivial) | ~2 minutes Suggested reviewers: ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Warning There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure. 🔧 ESLint
ESLint install timed out. The project may have too many dependencies for the sandbox. Comment |
|
This pull request has been ignored for the connected project Preview Branches by Supabase. |
BigSimmo
marked this pull request as ready for review
July 21, 2026 17:14
BigSimmo
enabled auto-merge (squash)
July 21, 2026 17:14
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 33cb0cd015
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (parity commit review) | 1aebf02 (fix landed a3b9a54) | clinical-governance-reviewer on the same commit: APPROVE-WITH-NITS. P2 (independently converged with the retrieval reviewer's P3): monitoring figure-escape lacked the dose-path multi-drug cross-entity guard — a bare wrong-drug schedule/level row in a multi-drug chunk could be entity-prefixed for a named-drug monitoring query; downstream gates verify text-vs-source presence, never attribution (worked lithium/valproate LFT path traced through finalize). FIXED in a3b9a54 exactly as its smallest-fix prescribed (guard at the :872-881 site now fires for monitoring_schedule; negative multi-drug test added, red-proven). Clean: unsupported figures impossible (admission-only change; promotion corpus guard + numeric verification + claim support all byte-unchanged); conservative failure intact (figure-bearing-only escape, schedule-free refusal pinned); adjacent-context safety held (sourceEvidenceText excludes adjacent_context; weeks refusal confirmed by probe); no PHI/provider/ranking surface. P3s: RAG impact line (present in the PR body — behaviour-change form, correct since the PR also carries the Option A retrieval change); multi-drug negative test (landed in a3b9a54). | Offline guard-chain trace + targeted vitest probes (named-drug guard, bare-figure admission, conservative gap, adjacent refusal). Provider/release gates deferred per confirmation boundary. | | ||
| | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (Option A rescue review) | 0abf3c9 | clinical-governance-reviewer on the S3 escalation rescue: APPROVE-WITH-NITS, no P0/P1. P2 = the mandated live canary pair itself (process gate, declared in the PR body; offline-green + review-approved proven insufficient for this surface 2026-07-20). P3s: multi-drug escalation-query recall edge (titled drug + untitled drug — fast-path return can skip the vector leg; recall limitation, not misattribution, mirrors the pre-existing allowlisted-class tradeoff); reviewer probe files must stay uncommitted (relocated to scratchpad). All six clinical concerns verified safe: wrong-document impossible (alias phrases must appear in the query; per-document alias groups, no cross-drug conflation), conservative availability (S3 purely additive via keyed-union merge; sibling retention test-pinned), live expansion acceptable (title-named correct-entity SOP in every firing shape), fail-closed double layer (adversarial short-circuit precedes the predicate; injection-forge case intent=protocol cannot fire — executed), governance metadata unbypassed (same attachDocumentRankingMetadata + status=indexed + access-scope filters), no PHI/provider/schema surface. | Reviewer checks: escalation-rescue suite 8/8, injection-forge intent derivation executed, static trace of the full S3 chain. Live canary pair = the trust gate, dispatched post-merge. | | ||
| | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (Option A wave verdict — no code change; #1040 merged as cde6c5c) | canary run 29827012719 (#61, main cde6c5c) vs banked #60 (29800029819) | OPTION A WAVE ADOPTED — FIRST FULLY-GREEN 44-CASE CANARY IN PROGRAM HISTORY (Blocking failures: None). (1) Option A payoff EXCEEDED: citation_failure_rate 0.0227→0; the neuroleptic-side-effect-escalation case flipped from wrong-doc→failed-generation→1-citation-fallback to **strong route, successful gpt-5.6-sol generation, passed in 15.4s with no fallback marker** — the rescued S3 retrieval fixed generation itself, not just the citation count; expected_source_hit 0.6364→0.6591. (2) Golden held exactly as the blast-radius analysis promised: 36/36 PASS, content_recall 1.0, mrr@10 0.8921 BYTE-IDENTICAL to the pre-wave baseline (zero ordering movement — no golden case fires the predicate), irrelevant@10 0.1083→0.0917 (slightly better). (3) Parity payoff PARTIAL: monitoring targeting 1/5→2/5 (olanzapine-lai flipped — previously called a retrieval-depth residual; quetiapine-dose also flipped on the dose side); lithium-range (232ch) + metabolic (73ch, byte-identical answer to #60) did NOT flip despite offline-proven fixes — their live chunk sets evidently contain no admissible schedule sentence even under the widened gate → reclassified as retrieval-depth/live-content residuals joining adhd; below the ≥3/5 target but strictly improved, no regression anywhere. Dose 2/5 vs 2/4: same passing count, applicable set grew (new quality-metformin-renal-dosing miss = eval-set churn, not regression). (4) No-worse EXCEEDED: relevance 0.5333→0.6 (the two-step watch-item slide FULLY REVERSED to the #58 level), targeting_rate 0.6667→0.6957, fail_closed 0.9 held, readability/artifact_leaks 1.0, route ceilings 0, grounded 1.0, unsupported_correct 1.0, numeric 0, p95 22.8s, red_result 3/3. Adoption per the measured-gain rule: primary goal achieved, three case flips, relevance recovered, zero regressions. Residual queue: monitoring retrieval-depth trio (lithium-range/metabolic/adhd), E-3d H2 discards, weekly ANSWER_CASE_LIMIT 8→44 raise now unblocked (gate would be green), comparison-class coverage. Wave spend +~$2-4 → Phase E + Option A total ~$12-20 of ≤$20. | Evidence: run #61 job log read (Threshold Status: None; Answer Metrics; neuroleptic diagnostics row; targeting metric_rates + 6-miss list; golden 36 PASS lines + summary). Revert drill NOT triggered. | | ||
| | 2026-07-21 | claude/clinical-kb-pwa-review-asi3wb (PR: I9 weekly coverage raise) | see PR head | ADDENDUM 5 post-green item I9 (plan-authorized "after reds fixed"): weekly scheduled canary ANSWER_CASE_LIMIT default 8→44 — the Sunday 18:00 UTC cron now guards the FULL answer-quality case set instead of the first 8 (both #57 blocking reds historically lived OUTSIDE the first 8, leaving the weekly gate blind to them). Unblocked by run #61 proving the citation gate green on the full 44. Cost: est +$1-2/week (user-authorized in the plan). Contract test pin updated in lockstep (eval-canary-workflow.test.ts). Dispatch shapes unchanged (input override still wins); operational-risk diff, plain-revert rollback. | check:github-actions PASS; check:ci-scope PASS; check:gate-manifest PASS (20/20); eval-canary-workflow contract 4/4; prettier clean; no provider calls | |
There was a problem hiding this comment.
When this PR branch advances or someone later reviews this scope, the ledger entry's see PR head value no longer identifies the reviewed commit, so the review-throttling check cannot compare the saved HEAD against the current HEAD and may either repeat work or skip based on ambiguous history. Please replace this placeholder with the concrete reviewed SHA for the I9 change, such as 6bc3ffbd5e1020dd1d0ba3211a4e8dbc0e18ca30.
AGENTS.md reference: AGENTS.md:L146-L146
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The final plan-authorized item of the Phase E / Option A program (ADDENDUM 5 I9, gated "after reds fixed"): the weekly scheduled canary's
ANSWER_CASE_LIMITdefault rises 8 → 44, so the Sunday 18:00 UTC cron guards the full answer-quality case set instead of the first 8. Both historical blocking reds (the neuroleptic citation failure and the route-ceiling breaches from baseline #57) lived outside the first 8 cases — the weekly gate was structurally blind to them. Canary run #61 (maincde6c5c, the merged #1040 wave) proved the citation gate green on the full 44 (Blocking failures: None — the first fully-green 44-case canary in the program), unblocking this raise.answer_case_limitinput still overrides the default.tests/eval-canary-workflow.test.ts) updated in lockstep.RAG impact: no retrieval behaviour change — workflow env default + contract test pin only; no
src/lib/rag/**edit.Verification
npm run check:github-actions— pin check passednpm run check:ci-scope— scope self-test passednpm run check:gate-manifest— 20/20 gates enforcedtests/eval-canary-workflow.test.ts— 4/4 (pin updated in lockstep)Risk and rollout
🤖 Generated with Claude Code
https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9
Generated by Claude Code
Summary by CodeRabbit
Chores
Documentation
Tests