Model Hygiene Review

Review divergence without assuming that either the accepted model or current implementation is automatically correct. Produce an advisory authority assessment before changing either side.

Workflow

1. Run the deterministic evidence collector for the requested stable ID. Omit the target only when the user requests every accepted component:

   just model-audit component.rtg.query

   Read the JSON bundle under build/model-audits/. Treat comparisons as leads, not findings.

2. Read references/authority-triage.md completely. Identify the model lifecycle, owner, requirements, satisfiers, realization binding, codecs, verification cases, and exact evidence nodes.
3. Inspect the public implementation protocol, behavior, black-box tests, adapters, and consumers. Do not infer a contract from private helpers.
4. Inspect git log --follow, the introducing commits on both sides, and any predecessor accepted specification. Determine whether code predates the model, implements a later decision, or merely drifted.
5. Invoke sysml-reference only when syntax or language semantics affect the conclusion. Report the specification, section, printed or physical page, and whether the source is normative or informative. Do not use training recall as specification evidence.
6. Classify every material difference using the authority-triage reference. Separate an explicit realization codec or intentional implementation freedom from public semantic drift.
7. Report compatibility impact, confidence, evidence on each side, proposed authority, and the smallest verification needed before remediation.

Safety Rule

This skill is review-only. Do not edit the model, implementation, tests, generated views, or evidence while performing the audit. If accepted behavior is unresolved, classify it as human_decision_required and pause for the human owner. Successful parsing, tests, or evidence reference resolution are evidence, not proof of semantic agreement or verification closure.

Output

For every finding report:

• stable component and affected model/implementation symbols;
• classification and confidence;
• model, implementation, consumer, history, and specification evidence as applicable;
• compatibility and state/invariant impact;
• proposed authority and remediation;
• verification needed and whether human approval is required.

Conclude with unresolved decisions and the exact commands used. Keep generated audit bundles ignored and advisory; do not add them to CI or commit them.