craft-critique: autoresearch re-run with conveyance evals (post-judgment-contract)

[sungjunlee]

Follow-up to Epic #145 / PRD-DP E3.1 — the first real-world exercise of the prescription-ratchet counterweight (craft-autoresearch/references/eval-guide.md § The prescription ratchet). Tracked as the stated next step in docs/status.md (craft-critique row).

#150 replaced craft-critique's fixed 5-section template with a judgment contract: convey (a) prioritized findings with severity + evidence, (b) strengths worth preserving, (c) fix ordering with rationale — shape scales with the artifact. The old eval suite scored section shape and is now void; the skill has no autoresearch evidence against its new contract.

Change
Design a conveyance-led suite and run one autoresearch pass:

• Conveyance binaries (floor checks): every [HIGH]/[MED] finding carries file:line or quoted-phrase evidence; strengths name specific elements, not generic positives; fix ordering states a rationale (dependency/reach/risk/reversibility); when failure outputs are supplied, each [HIGH] ties to one.
• Comparative (leads KEEP/DISCARD): blind A/B on actionability — could a reader start fixing without re-reading the artifact?
• Proportionality: a small artifact must get a short critique — no fixed section scoring anywhere in the suite.
• ≥6 inputs with train/holdout split per the E1.1 discipline; run artifacts under ~/.craftkit/autoresearch/craft-critique/.

AC

• Suite contains zero section-shape assertions; KEEP/DISCARD is led by outcome/comparative evals
• Baseline + at least one mutation cycle run against the judgment contract; kept mutations (if any) land with finding prose in commit bodies
• docs/status.md craft-critique row updated with the new evidence pointer
• If loop pressure pushes toward re-adding fixed format rules, record it as a ratchet observation in eval-guide § The prescription ratchet instead of applying it

[coderabbitai]

🔗 Related PRs

#3 - Add skill-radar curation layer + craft-skill-spec meta-skill [merged]
#5 - craft-handoff: pair per-session prompt with per-project rich doc [merged]
#34 - [codex] document reviewable goal guidance [closed]
#67 - [codex] Keep CraftKit boundaries lightweight [closed]
#79 - docs(readme): clarify contract evidence status [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[sungjunlee]

Closing as out of scope for #169. A conveyance-eval autoresearch re-run is the opposite of this pass (more eval ceremony on a skill that already moved to a judgment contract). craft-critique stays as-is unless real misuse shows up.