Mostly efficient and instruction-oriented, but includes conceptual padding (e.g., 'These are NOT intentional deception — they are failure modes of optimizing agents') and a 'Why This Exists'/'Acknowledgements' rationale that assumes less competence than warranted; could be tightened.

Provides concrete file-scan glob patterns, exact reviewer MCP call signatures with model/config/sandbox parameters, a full copy-paste audit prompt, and exact output report templates — fully executable guidance.

A clearly sequenced 4-step workflow (collect artifacts → send to reviewer → parse/write report → print summary) with explicit validation checkpoints (reviewer reads directly, PASS/WARN/FAIL verdict) and an integration diagram.

Structure is present and references to shared-references/*.md are signaled, but the body is somewhat monolithic with the ~70-line audit prompt and full report templates inline; the referenced shared-references files do not exist in the bundle, so navigation is not actually resolvable.

Lists multiple concrete actions — 'check for fake ground truth, score normalization fraud, phantom results, and insufficient scope' — matching the 'lists multiple specific concrete actions' anchor.

Explicitly answers both what ('Audit experiment integrity... Uses cross-model review to check for...') and when ('Use when user says...'), with explicit 'Use when' trigger guidance.

Provides good coverage of natural terms users would say in both English ('check experiment integrity', 'audit results') and Chinese ('审计实验', '实验诚实度'), matching the top anchor.

A clearly defined niche (experiment integrity auditing) with distinct bilingual triggers unlikely to fire for unrelated skills, matching the 'clear niche with distinct triggers' anchor.