Judge Autoform work as an agent

There is no autoform on PATH. Run it from the project under review as uv run --project "<AUTOFORM_PLUGIN_ROOT>" autoform, where the root is the absolute plugin directory holding skills/ and pyproject.toml; give a subagent that prefix.

• For a roadmap or blueprint, read roadmap quality, inspect its declared sources and coverage boundary, and validate its DAG.
• For Lean code, read faithfulness, proof integrity, code quality, and Mathlib style; gather each one's evidence.

Keep objective evidence separate from judgment. Never claim compilation, declaration resolution, axiom cleanliness, source coverage, or dependency correctness without showing how it was checked. If required sources are absent, return insufficient evidence rather than guessing.

Regenerate skeleton evidence from the exact candidate after its Lean build. Do that only in a trusted checkout or an operating-system sandbox: the command evaluates Lake configuration and project Lean metaprograms, and its resource bounds are not a security boundary. Candidate code runs during extraction and can forge process output, so treat its report as advisory when the checkout is not trusted. Treat a stale-build refusal as insufficient evidence; never approve a current source excerpt paired with an older compiled declaration. Record the skeleton hash, the evidence hash of the packet that was read, and, for a source-faithfulness verdict, the article review hash; the skeleton reference says what each binds. These hashes are provenance evidence, not reviewer authentication or an approval key. If the user has paused skeleton evidence during proof development, skip it and Formalize's import-edit baselines, say so, and record none of these hashes; the rubrics' evidence still applies, and so do autoform check and audit with --lean-root.

Report findings first, ordered by severity and tied to files or nodes. Then give the rubric scores, weighted verdict, commands run, unresolved questions, and a short remediation list. Do not edit the reviewed work unless the user separately asks for fixes.

Score faithfulness and integrity separately: a sorry-free proof that drops a source hypothesis still fails faithfulness; report it, do not average it away.