aat-reviewer

name	aat-reviewer
description	Review ArchSig AAT Observable Bundle artifacts and translate AAT concept evidence into bounded design review actions. Use when Codex is asked to review AAT-facing ArchSig artifacts, explain invariant/witness/signature/operation/boundary evidence, or decide next evidence for architecture review.

AAT Reviewer

Purpose

Read aat-observable-bundle/v0.5.0 and related ArchSig artifacts as AAT-facing review evidence. This skill turns deterministic evidence into review cues while preserving the boundary between tooling evidence, LLM judgment, human review, and Lean proof.

Inputs To Prefer

• aat-observable-bundle/v0.5.0 and validation report
• AIR and AIR validation
• ArchMap observation artifact with canonical ArchSig input-validation evidence
• theorem-check report
• Feature Extension Report
• obstruction-witness/v0.5.0
• PR quality analysis and policy decision reports

Workflow

1. Validate the bundle before using it.

${ARCHSIG_BIN:-archsig} aat-observable-bundle \
  --input .archsig/aat/aat-observable-bundle.json \
  --out .archsig/aat/aat-observable-bundle-validation.json

2. Read the AAT concept coverage.

   • Check conceptMappings for ArchitectureObject / ComponentUniverse, ObstructionWitness, TheoremBoundary / NonConclusion, ArchitectureOperation, Projection / Observation / LSP / DIP, FeatureExtension, semantic diagram, state/effect law, repair/synthesis, and analytic axis.
   • Treat out-of-scope concepts as explicit boundaries, not missing implementation failures.

3. Separate evidence states.

   • measuredZero means measured zero only inside the selected universe.
   • measuredNonzero means selected evidence found a witness or nonzero axis.
   • unmeasured, outOfScope, private, unavailable, unsupported, and dynamicBlindSpot are not measured zero.

4. Translate nonConclusions.

   • Evidence gap: ask for the next artifact or measurement.
   • Non-finding: state that no selected evidence was found without claiming global absence.
   • Blocked formal claim: do not promote tooling evidence to Lean proof.
   • Review guardrail: keep semantic, runtime, static, and analytic surfaces separate.
   • Next evidence: name the smallest artifact that would improve the review.

5. Produce bounded review output.

   • Tie every finding to sourceRefs, witnessCatalog, operationCandidates, theoremBoundaries, or reviewActions.
   • Use llmReviewSurface.reviewQuestions as the question template.
   • Leave risk acceptance, product decision, and merge approval to human review.

Output Shape

AAT review state
- ...

Measured witnesses
- ...

Evidence gaps and guardrails
- ...

Operation / repair interpretation
- ...

Next evidence
- ...

Non-conclusions
- ...

Use Japanese for user-facing reports in this repository, unless the user asks otherwise.