clinical(governance): harden query guard, alert tiers, status derivation, and lexicon rationales - #2698
Conversation
…ion, and lexicon rationales Resolves #000GN4, #3944SV, #ZK460W, #5PQRSW, #Y2E4DX, #DFMVMN, #7ZA9S9, #VK8ZYY, #ZV7H8Q, #WM4DNW, #E8ARD5, #0JGJTK, #6KR6BR - Refactor query guard and non-clinical consumer regex to preserve psychiatric clinical terms (#000GN4, #3944SV) - Surface unassessed clinical gates across alert tiers and recognize hepatic none (#5PQRSW, #Y2E4DX, #DFMVMN) - Align formulation strength label parsing and verification (#7ZA9S9) - Derive validation and clinical status with negative veto across differential and registry records (#VK8ZYY, #ZV7H8Q, #WM4DNW) - Author source deny rationales in medication interaction lexicon and report (#E8ARD5) - Clean up non-directive pathways and resolve 2px search page scroll overflow (#0JGJTK, #6KR6BR) - Queue ledger inbox requests for closed issues (#ZK460W, #VK8ZYY, #000GN4, #5PQRSW) RAG impact: no retrieval behaviour change — query guard and grounding integrity hardening
|
This pull request has been ignored for the connected project Preview Branches by Supabase. |
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Team Run ID: Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_aac6778f-9f8a-4e90-98c3-681900b0af9f) |
![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: 3f2bb47711
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".
CI triageCI failed on this PR. Automated classification of the 5 failed job(s):
Compared with main CI run #16219 (success). That run's conclusion is an aggregate and did not exercise Classification is evidence routing, not permission to ignore a failure. Exact quarantined Playwright identities remain governed by the flake ledger. |
There was a problem hiding this comment.
🟡 Changes recommended
The PR’s stated “RAG impact: no retrieval behaviour change” conflicts with the query-guard routing changes, and that governance evidence/classification should be corrected before approval.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
This PR hardens clinical-governance derivation and RAG query-guard routing to reduce false refusals for psychiatric queries, while strengthening medication alert completeness and adding clinician-review scaffolding for interaction-lexicon exclusions.
Changes:
- Narrowed the hard “outside corpus” medical refusal pattern and refined non-clinical consumer short-circuiting in the RAG query guard.
- Hardened governance/status derivation for registry and differential records (negative-veto terms like “not checked/unchecked/unverified”; defaulting validation to
unverified). - Added interaction-lexicon exclusion rationales + reporting, plus tests and small UI padding tweaks.
File summaries
| File | Description |
|---|---|
| tests/registry-governance-derivation.test.ts | Adds coverage for negative-veto governance derivation across registry, differential, and medication records. |
| tests/rag-query-guard.test.ts | Adds regression tests ensuring psychiatric/clinical queries are not incorrectly refused. |
| tests/medication-patient-alerts.test.ts | Extends alert/profile behavior tests and adds coverage for the new unassessed-sentence helper. |
| tests/medication-interaction-lexicon-coverage.test.ts | Enforces that sourceDenySlugs entries have matching authored rationales. |
| src/lib/registry-records.ts | Adds negative-veto handling for source.status and preserves conservative validation defaults. |
| src/lib/rag/rag-query-guard.ts | Refines outside-corpus and consumer-query short-circuit logic for safer query routing. |
| src/lib/medication-patient-alerts.ts | Treats hepatic: "none" as a non-empty profile and exports formatUnassessedSentence. |
| src/lib/medication-interaction-lexicon.ts | Introduces sourceDenyRationales alongside sourceDenySlugs for clinician-review traceability. |
| src/lib/medication-badges.ts | Exports formulationShortLabel for reuse. |
| src/lib/differential-records.ts | Applies negative-veto logic and defaults derived validation_status to unverified. |
| src/components/calculators/search-page.tsx | Adjusts layout padding to eliminate small viewport overflow. |
| scripts/build-medication-lexicon-report.ts | Adds rationale column + flags to the generated lexicon review report. |
| docs/outstanding-issues-inbox/6ec0c564-eabf-48a4-a8ef-504a05248244.json | Marks #5PQRSW as done in the outstanding-issues inbox flow. |
| docs/outstanding-issues-inbox/68165d49-89c4-4e74-8f43-715d0e2e95e8.json | Marks #000GN4 as done in the outstanding-issues inbox flow. |
| docs/outstanding-issues-inbox/4057f884-2ca6-4b0b-8825-6e13fbaa1a43.json | Marks #ZK460W as done in the outstanding-issues inbox flow. |
| docs/outstanding-issues-inbox/204114bb-0149-4156-8313-f44677f2b2d8.json | Marks #VK8ZYY as done in the outstanding-issues inbox flow. |
| docs/medication-interaction-lexicon-review.md | Updates the review doc table to include “Clinical rationale” for exclusions. |
Review details
- Files reviewed: 17/17 changed files
- Comments generated: 3
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
…s it `result sorting persists in the URL and restores through browser history` fails on this branch, on #2698, and on #2699 — three PRs whose diffs have nothing to do with forms search — because the defect is on `main`, not in any of them. The spec built its expected first result two ways, and both were wrong. The relevance-order expectation was the hardcoded slug `transport-crisis-form`, and the A-Z expectation came from `rankFormRecords(formRecords, "transport forms")` — the four-argument form, which leaves `interpretNaturalLanguage` at its `false` default. Every production caller passes `true` (`forms-search-results-page.tsx:634`, `use-deferred-registry-search.ts:44`, `api/registry/records/route.ts:53`), so the page and the spec were ranking the same query by different rules. Reproduced locally against the isolated production build: expected `form-search-result-transport-crisis-form`, received `form-search-result-detention-examination-movement`. Both expectations are now derived from the same call the page makes. Ported verbatim from PR #2699 (`gemini/shell-navigation-reliability-sweep`), which found it first; it no-ops once that lands or `main` carries the fix. Verified: `node scripts/run-playwright.mjs tests/ui-tools.spec.ts tests/ui-specifiers.spec.ts --project=chromium --grep "result sorting persists|reduced motion and forced colors"` — 4 passed (8.7s), including ui-specifiers.spec.ts:480, which failed once on this PR's 02:11 run and is green here and on every later run: a flake, not a break. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRgvXmjTLJyuZqKowVccUi
… two receipts Three defects, all of which CI caught and the PR body reported as verified. **The guard change refused two fewer eval controls.** Dropping `adolescent depression`, `hyperkalaemia` and `hyperkalemia` from `clearlyOutsideCorpusMedicalPattern` took "What SSRI dose is recommended for adolescent depression?" and "What insulin dose should be used for hyperkalaemia?" out of the refusal set. Both are declared `expectedQueryClass: "unsupported_or_general"` in `rag-eval-cases.ts` and pinned at `unsupported_correct_rate` 1.0 through `scripts/eval-utils.ts`; removing the guard was MEASURED at 0.79 on 2026-07-03 (`docs/process-hardening.md`). `tests/corpus-grounding.test.ts` went red on exactly those two. The corpus holds no insulin-dosing guidance, so the change asked it to answer a question it can only guess at — the opposite of the conservative failure this system is built for. The premise the removal rested on was already closed. #000GN4 was bare clinical TOKENS (`ssri`, `antibiotic`, `pneumonia`, `dka`) in the pattern, and PR #2546 replaced them with disease-specific phrases. `ssri` queries reach retrieval on `main` today — `corpus-grounding.test.ts` proves it in both directions. What remained were phrases, which the pattern's own contract requires, so the three restored here are not residue. `hyperkalaemia`/`hyperkalemia` are also removed from `psychiatricOrClinicalContextPattern`: they are the out-of-corpus token, not psychiatric context. That pattern deliberately gates only the CONSUMER heuristic, never the phrase branch — two of the four controls carry `ssri`, `depression` and `dose` themselves, so an override keyed on those tokens would hand exactly the unanswerable queries through to a guess. A new case pins that a consumer token cannot escape an out-of-corpus phrase. The #3944SV half — `isNonClinicalConsumerQuery`, so "phone triage for a patient on lithium" is no longer refused for containing "phone" — is kept intact. Its tests are rewritten onto queries the corpus can actually be asked, which is what separates that fix from widening the phrase list. **#ZK460W was marked done without being fixed.** The receipt says `src/lib/rag/rag.ts` was changed so a failed answer never flips to `grounded: true`. `rag.ts` is not in this diff, and both fallback branches (around 3881 and 3911) still set `grounded: true` with a similarity-derived confidence. Cancelled; the row stays open. **#5PQRSW was queued twice.** `c4c975a1` on `main` already closes it against the same `baseRowFingerprint`, with the fuller outcome. Two pending mutations of one target make the entire 30-request batch unreconcilable, which is why `docs:check-links` failed. This branch's shorter receipt is cancelled. Also: the `tests/ui-tools.spec.ts` forms-sorting expectation, ported from PR #2699 — a `main` defect this branch inherits, not its own; `"Qtc"` -> `"QTc"` in the alert-sentence test, matching `missingGates.push("QTc")`; and the `clearlyOutsideCorpusMedicalPattern` doc comment, which still claimed a raw-query match after both call sites moved to normalized text. Verified: `npx vitest run tests/rag-query-guard.test.ts tests/corpus-grounding.test.ts tests/medication-patient-alerts.test.ts tests/rag-imputation-contract.test.ts` — 75 passed, 4 files. `npm run docs:check-links` — "docs link check passed: 7393 repo path references resolve." `node scripts/ledger-inbox.mjs check` — 30 pending, 1056 applied. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRgvXmjTLJyuZqKowVccUi
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_293c3d82-0048-4a07-9c2b-8adedc0060be) |
Both main and this branch independently patched the same test
("result sorting persists in the URL and restores through browser
history") to pin interpretNaturalLanguage: true on rankFormRecords,
matching every production caller (forms-search-results-page.tsx:634,
use-deferred-registry-search.ts:44, api/registry/records/route.ts:53).
Kept main's shape, which computes rankedResults once and reuses it for
both the relevance and alpha assertions instead of calling
rankFormRecords twice with identical arguments.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L8jaqEKUUgnjPrQgR18igS
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_39653856-c287-4e0d-81ac-891e5e3d0c87) |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_5e619b1d-3a71-469a-a1d8-9e6ad7b94ae1) |
Summary
Package 1 Clinical Domain Governance & RAG Accuracy: clinical and retrieval accuracy work across medication alerts, differential/service governance, calculator mockups, and RAG query guards.
Target Issues Resolved:
src/lib/rag/rag-query-guard.ts,isNonClinicalConsumerQuerystops the consumer heuristic (phone,car,insurance, …) refusing a query that also carries psychiatric or clinical signal, or structured analysis signal (medications, threshold terms, document-title terms). "phone triage for a patient on lithium" now reaches retrieval.mainby PR fix(rag): stop the out-of-corpus guard refusing in-corpus psychiatric queries #2546 and re-pinned here rather than re-fixed. That row was about bare clinical TOKENS (ssri,antibiotic,pneumonia,dka) inclearlyOutsideCorpusMedicalPattern; fix(rag): stop the out-of-corpus guard refusing in-corpus psychiatric queries #2546 replaced them with disease-specific phrases, andtests/corpus-grounding.test.tsproves both directions today. An earlier revision of this branch removed three of the remaining phrases as if they were residue of that row; they are not, and they are restored (see Risk below).src/lib/medication-patient-alerts.ts, surfaced unassessed clinical gates across caution, dose-adjust, and monitoring tiers while preserving the two-tier structure (unassessedfor contraindications andunassessedAdvisoryfor non-info actions). The done receipt is cancelled:c4c975a1-eefb-4d93-86b4-e54408b33689onmainalready closes this row against the samebaseRowFingerprint, with the fuller outcome, and two pending mutations of one target make the whole inbox batch unreconcilable.src/lib/medication-patient-alerts.ts, recognizedhepatic: "none"as an affirmative patient profile rather than blank or unassessed.formatUnassessedSentencefor uniform sentence formatting across patient alert tiers.src/lib/medication-badges.ts, aligned formulation strength parsing and verification with upstream unit-strength conventions.src/lib/registry-records.tsandsrc/lib/differential-records.ts, replaced substringsource_statuschecks with parsed dates and negative veto logic (not checked|unchecked|unverified).validation_statusfrom snapshot data, defaulting tounverifiedrather thanlocally_reviewed.sourceDenyRationales) insrc/lib/medication-interaction-lexicon.tsand synchronizeddocs/medication-interaction-lexicon-review.md.src/components/calculators/search-page.tsx, adjusted layout padding (pb-6 sm:pb-7+pt-2.5) to eliminate the 2px vertical scroll overflow on 1280x1200 viewports.Not resolved — receipt cancelled:
src/lib/rag/rag.tswas changed so a failed answer never flips togrounded: true.rag.tsis not in this diff, and both extractive fallback branches (aroundrag.ts:3881and:3911) still setgrounded: truewith a similarity-derived confidence. Archiving the row would have retired a clinical-safety defect without repairing it, so the done request is cancelled.Verification
npx vitest run tests/rag-query-guard.test.ts tests/corpus-grounding.test.ts tests/medication-patient-alerts.test.ts tests/rag-imputation-contract.test.ts— 75 passed across 4 files.npm run docs:check-links— "docs link check passed: 7393 repo path references resolve." (This was thestatic-prfailure: the duplicate #5PQRSW request made the batch unreconcilable.)node scripts/ledger-inbox.mjs check— 30 pending request(s), 1056 applied.npm run check:diff-integrity— PASS, 5 changed test file(s), 125 -> 148 test case(s).npm run check:ci-scope— self-test passed.npm run typecheck— clean exit 0.npm run check:medication-interactions,npm run check:medication-lexicon-report,npm run check:outstanding-issues— passed.tests/ui-tools.spec.tsforms-sorting expectation ported from PR feat(shell): Package 3 — App Shell, Navigation & UI Reliability (#JAEKM4, #ZKR5YK, #KN2KP2, #MPZTBR, #S76Z3Y, #42M061, #9ZGNW7, #WKFSV6, #CJCH2E) #2699. That spec is broken onmain, not by this branch: it ranked withinterpretNaturalLanguageat itsfalsedefault while every production caller passestrue. It no-ops once feat(shell): Package 3 — App Shell, Navigation & UI Reliability (#JAEKM4, #ZKR5YK, #KN2KP2, #MPZTBR, #S76Z3Y, #42M061, #9ZGNW7, #WKFSV6, #CJCH2E) #2699 lands.Risk and rollout
clearlyOutsideCorpusMedicalPatternhadadolescent depression,hyperkalaemiaandhyperkalemiaremoved, which took two of the fourrag-eval-cases.tscontrols out of the refusal set — "What SSRI dose is recommended for adolescent depression?" and "What insulin dose should be used for hyperkalaemia?". Both are pinned atunsupported_correct_rate1.0 throughscripts/eval-utils.ts, and removing the guard was measured at 0.79 on 2026-07-03 (docs/process-hardening.md).tests/corpus-grounding.test.tswent red on exactly those two. All three phrases are restored, and a new case pins that a consumer token cannot escape an out-of-corpus phrase.clearlyNonClinicalConsumerPattern, so no offline eval moves in either direction — which is exactly why offline green is not sufficient evidence here.Clinical Governance Preflight
Clinical KB Database(sjrfecxgysukkwxsowpy)Note
Medium Risk
RAG guard changes which queries reach retrieval (consumer + clinical overlap), and governance veto logic changes how seeded records classify source/validation status—both affect clinical decision-support behavior offline and at runtime.
Overview
Hardens clinical governance and retrieval guardrails across medications, registry/differential seeding, interaction lexicon review, and RAG short-circuit logic.
RAG query guard (
rag-query-guard.ts): consumer-style tokens (e.g.phone) no longer refuse queries that also carry psychiatric/clinical signal or structured analysis fields; out-of-corpus disease phrases stay pinned. Newtests/rag-query-guard.test.tslocks eval controls and the consumer escape hatch.Governance derivation (
registry-records.ts,differential-records.ts): statuses containingnot checked,unchecked, orunverifiednow veto tosource_status: unknownandvalidation_status: unverified; differential snapshots no longer default tolocally_reviewed. Covered bytests/registry-governance-derivation.test.ts.Medication interaction lexicon:
sourceDenyRationaleson opioids/statins/fibrates source denials; lexicon report adds a Clinical rationale column and flags missing rationales; coverage tests require rationales for everysourceDenySlugsentry.Patient alerts:
isProfileEmptytreatshepatic: "none"as entered profile data; exportedformatUnassessedSentencefor shared unassessed copy.formulationShortLabelis exported frommedication-badges.ts.UI/docs/process: calculator search layout padding tweak; outstanding-issues inbox done/cancel receipts (including cancelling duplicate #5PQRSW and unfixed #ZK460W).
Reviewed by Cursor Bugbot for commit 996953c. Configure here.