Summary

First Phase E delivery (user-approved with budget): give the canary an opt-in answer_quality_eval dispatch input that runs the 30-case eval:answer-quality fixture — five answer-quality metrics (relevance, readability, artifact leaks, intent coverage, fail-closed) plus the per-intent structural targeting metric (dose→figure, red-result→withhold action, monitoring→schedule). E-1 recon found this instrument has never run in CI, and the canary's answer gate samples only 8 of the 44 eval:quality cases — so the Phase E baseline needs this step.

Design constraints honored:

• Dispatch-only and default-off: scheduled runs and every existing dispatch shape are byte-identical (if: workflow_dispatch && input == 'true'); follows the feat(ci): eval-canary dispatch inputs — staged ranking-config override + branch evals (ADDENDUM 4 B-PR-1 / D-1) #1003 input-only precedent.
• Informational by construction: the script exits 0 without --targeting-floor, so this step can never fail a canary — quality gates stay owned by the eval:quality step.
• The tee'd log lands in .local/eval-canary/ and rides the existing pinned artifact upload; 45-minute job timeout comfortably absorbs the ~10-minute step.

Next (after merge): the E-2 baseline dispatch — answer_case_limit=44 + answer_quality_eval=true — one run giving the full answer-quality baseline on both instruments plus the golden retrieval eval as the built-in no-regression net. Estimated $3–8 of the user-authorized ≤$20 Phase E envelope; reported before/after dispatch.

RAG impact: no retrieval behaviour change — dispatch-input-only workflow addition; retrieval steps, schedule, and defaults untouched.

Verification

• npm run check:github-actions — pass (no new actions; existing pins untouched)
• npm run check:ci-scope — pass
• npm run check:gate-manifest — pass
• npx prettier --check on both changed files — clean
• No provider calls in this PR (workflow-definition change only); the paid dispatch happens separately under the approved envelope

Risk and rollout

• Risk: low — a conditioned, default-off step in a dispatch-only path; scheduled behaviour provably unchanged (condition requires workflow_dispatch AND explicit 'true'). The step is informational and cannot fail the run.
• Rollback: revert.
• Provider or production effects: none from this diff. When the input is enabled on a dispatch, the step spends 30 live generation cases ($1–3) — always under an explicit user-approved dispatch.

Clinical Governance Preflight

• Source-backed claims still require linked source verification before clinical use — no behaviour change
• No patient-identifiable document workflow was introduced or expanded
• Supabase target remains Clinical KB Database (sjrfecxgysukkwxsowpy) — same env block, no new secret exposure
• 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
• Deployment classification/TGA SaMD impact was checked: no clinical decision-support behaviour changed — measurement only

🤖 Generated with Claude Code

https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9

Generated by Claude Code

Summary by CodeRabbit

• New Features

  • Added an optional dispatch-only workflow input to run an additional 30-case answer-quality/targeting evaluation.
  • When enabled, the workflow produces a dedicated log artifact containing the targeting evaluation output.

• Documentation

  • Updated the Branch Review Ledger with the Phase E kickoff details, including the default-off behavior, canary sampling notes, and baseline follow-up parameters.