Imported from visionary-5/knotrail (skills/audit-performance-policy/SKILL.md). Install upstream with npx skills add visionary-5/knotrail --skill audit-performance-policy. Copyright stays with the author.

Audit Performance Policy

Produce a release verdict from deterministic evidence. Never infer performance from prose, screenshots, or an LLM judgment.

This is a domain reference Skill showing how Knotrail's reusable trust primitives can govern a deterministic GPU performance workflow. It is not the Knotrail product identity or a required repository-research stage.

Contract version

• Skill release: 1.0.0.
• Accepts: immutable performance corpus analysis plus named gate policy.
• Emits: knotrail-policy-audit-v1.

Capabilities

• Deterministic correctness and oracle-regret gate evaluation.
• Immutable input hashing and reproducible run IDs.
• Machine-readable verdicts and standalone evidence dashboards.
• Release blocking with explicit human approval separation.

Workflow

1. Confirm the input is a JSON artifact with base_analysis.gate, base_analysis.suite, and per-graph evidence.

2. Read references/contract.md before integrating a new producer or changing a gate.

3. Run:

   python3 scripts/run_audit.py \
     --input <corpus-analysis.json> \
     --json-output <audit.json> \
     --html-output <audit.html> \
     --policy <policy-name>

4. Treat verdict.status=rejected as final for the current candidate. Do not ask another agent or a human to override a failed deterministic gate.

5. When the verdict is eligible-for-human-review, request explicit human approval. Never interpret gate passage as release authorization.

6. Attach the JSON report, HTML dashboard, source SHA-256, and run ID to the AgentTeams task state.

Safety boundaries

• Run this Skill from the Independent Verifier identity.
• Keep policy-writing and sealed-evaluation permissions separate.
• Do not modify source artifacts, selector code, corpus membership, or gate thresholds.
• Do not expose paper-only untouched data to the Policy Engineer.
• On malformed input, missing fields, hash mismatch, or tool failure, return a failed task with no release decision.
• Re-running the same immutable input must produce the same run ID and verdict.

Outputs

• knotrail-policy-audit-v1 JSON evidence.
• A standalone HTML evidence dashboard.
• No code changes and no release side effects.