write-review-artifact

Write Review Artifact

Output schema

---
agent: technical-reviewer
version: 1.0
draft_path: corpus/drafts/{slug}.md
draft_sha256: {hash}
reviewed_at: ISO8601
topic_hint: {string or null}
total_claims: N
counts:
  simplified-correct: N
  simplified-boundary: N
  wrong: N
  contested: N
  overclaim: N
blockers: N                              # count of wrong claims
glossary_alignment_notes: N
go_no_go: GO | GO-WITH-HEDGES | NO-GO
timebox_used_hours: N.N
---

Body sections (exact order, no exceptions)

1. ## Summary — one paragraph, plain English, counts + verdict.
2. ## Blockers — every wrong claim, numbered, excerpt + why-wrong + primary source + suggested fix. Empty section if zero.
3. ## Claims — full per-claim table: #, excerpt, location, classification, confidence, primary_source, action, note.
4. ## Boundary-Break Suggestions — for every simplified-boundary.
5. ## Glossary Alignment — terms where writer's glossary diverges from field. Empty if zero.
6. ## Could-Not-Verify — claims where primary source was paywalled / unreachable / not findable.
7. ## Research Log — queries run, sources fetched, time spent. Audit trail.

Go/no-go rule (mechanical)

• blockers = count(wrong)
• If blockers > 0 → NO-GO.
• Else if contested + overclaim > 0 → GO-WITH-HEDGES.
• Else → GO.

Guardrails

1. Never emit GO if blockers > 0.
2. Never omit Could-Not-Verify section, even empty — an empty section is the positive signal.
3. Never modify the draft. Only write to the review path.
4. draft_sha256 recorded to detect re-runs on a changed draft.
5. Body sections in the exact order above. No interleaving.
6. Overwrite if re-run on the same date for the same slug; append -v2 suffix if running a second time after a revision.