artifact-review

Artifact Review

Don't ask permission. Just do it.

This is a release-readiness review skill. It does not invent new claims or run new experiments. It checks whether the current artifacts are safe to share.

Required Outputs

• review/artifact_review.md
• review/release_checklist.md
• review/release_gate.json

Review Scope

Use this for any mix of:

• paper/draft.md
• paper/figures_manifest.md
• review/draft.md
• experiment_res.md
• figure bundles
• README.md
• docs/index.html

Review the artifact set in one or more of these modes:

• paper review
  • checks claim scope, evidence binding, baseline wording, and abstract/results discipline
• figure review
  • checks units, legends, captions, readability, and evidence labels
• release page review
  • checks first-screen clarity, artifact entry points, and scope-boundary wording
• style review
  • checks paragraph discipline, quantitative grounding, adjective inflation, and result-vs-interpretation separation

Workflow

Step 1: Inventory the Artifact Set

List the files being reviewed, the headline claims they appear to make, the source artifact path for each headline claim when available, which figures or tables support them, and which review mode applies to each file (paper review, figure review, release page review, or style review).

Step 2: Review Findings First

Write artifact_review.md as a findings-first review using severity levels:

• P0 = unsafe to publish as-is
• P1 = materially weakens the claim or readability
• P2 = polish or consistency issue

Each finding must include:

• the problem
• the affected file(s)
• the evidence_path (N/A if the issue is structural rather than evidence-bound)
• the affected_claim_id (N/A if the issue is not tied to a specific claim)
• why it matters
• the concrete fix

Also write a top-level line:

release_verdict: HOLD | CONDITIONAL_GO | GO

Use these verdict rules:

• HOLD
  • any P0 finding exists
  • a headline metric has no baseline, no protocol/guardrail, or no source artifact
  • simulator/proxy evidence is written as runtime evidence
• CONDITIONAL_GO
  • no P0 findings exist, but one or more unresolved P1 findings remain
• GO
  • no P0 findings remain
  • no unresolved P1 finding weakens a headline claim
  • every headline claim can be traced to a concrete source artifact

Step 3: Check Release Readiness

Write release_checklist.md using the checklist in references/review-checklist.md.

If style review applies, also use references/style-review-checklist.md. Then write review/release_gate.json using references/release-gate-template.md.

If a paper-facing figure set exists, explicitly check the figure-text contract across:

• paper/claim_inventory.md
• paper/figures_manifest.md
• the first prose callout
• the figure caption
• the LaTeX or Markdown figure block

release_gate.json should include:

• release_verdict
• generated_at
• review_scope
• blocking_findings
• p1_findings
• checked_files
• stale_if_any_newer_than

Use stale_if_any_newer_than to list the release-facing artifacts that would invalidate the current gate if they change later, for example:

• paper/draft.md
• paper/claim_inventory.md
• paper/figures_manifest.md
• README.md
• docs/index.html

Required Checks

1. Every headline metric has a baseline, protocol/guardrail, and source artifact.
2. Simulator/proxy evidence is not written as runtime evidence.
3. Figures have readable titles, units, legends, and captions.
4. The first screen of README/docs answers:
   • what this is
   • how to use it
   • what artifacts exist
   • what the scope boundary is
5. Unsupported claims are downgraded or explicitly marked as open.
6. Results paragraphs are quantitative and baseline-anchored.
7. Conclusion and abstract do not introduce claims that exceed the allowed confidence or section scope.
8. Every headline claim has a matching figure, table, or explicit text-only justification.
9. Every paper-facing figure has supports_claim_ids, a usable callout_sentence, and a caption that names baseline, metric, evidence type, and protocol when relevant.
10. Figures are introduced before or adjacent to the claims they are supposed to support.
11. review/release_gate.json matches the current verdict and names the files that would make the gate stale if changed later.

Safety Rules

1. If the evidence trail is broken, flag it. Do not repair it with guesswork.
2. Prefer short, specific findings over generic writing advice.
3. Review the artifact that exists, not the artifact you wish existed.
4. If a sentence sounds impressive but is not measurable, downgrade it or flag it.