Conciseness

The body is mostly efficient (exact reviewer prompt, report templates, checklist), but includes unnecessary material: the manual-review calling convention and config block appear twice nearly verbatim (Reviewer Calling Convention section and Step 2), the opening meta-blockquotes reference undefined concepts ('this slot's tier', 'source/image inputs and deterministic gates'), and the Integration section documents other skills' (/result-to-claim, /paper-write) internal behavior. Fits anchor 3 ('could be tightened') rather than 2, since no section is pure padding.

Actionability

Fully executable guidance: a complete copy-paste-ready reviewer prompt with placeholder slots, exact tool names and configs (mcp__codex__codex with 'model_reasoning_effort': 'ultra', sandbox read-only), the full manual-review config, and complete EXPERIMENT_AUDIT.md/.json templates plus a printed-summary format. Matches anchor 5.

Workflow Clarity

Four clearly sequenced, role-labeled steps (Collect Artifacts → Send to Reviewer → Parse/Write Report → Print Summary) with validation gates (REVIEW_UNAVAILABLE on missing/same-family reviewer identity, stop-and-print-install-command when the manual MCP is unavailable) and PASS/WARN/FAIL verdict branching. Not 5 because the reviewer-identity requirement blockquote is misplaced inside Step 1 (artifact collection) and the dense header routing note muddies the workflow entry point.

Progressive Disclosure

The body is well-sectioned with clearly signaled one-level-deep references (shared-references/reviewer-routing.md, review-tracing.md, integration-contract.md, external-cadence.md), and inlining the ~70-line audit prompt is justified since it must be passed verbatim. No bundle files exist (references/, scripts/, assets/ absent) and everything lives in one ~320-line file, so it does not reach anchor 5's clean overview-plus-split structure; minor organization gaps keep it at 4.

Specificity

The description lists multiple specific concrete actions — 'check for fake ground truth, score normalization fraud, phantom results, and insufficient scope' — plus the mechanism ('Uses cross-model review (external reviewer backend)'), giving comprehensive coverage of the skill's checks. It is not 4 because there are no minor gaps in the enumerated capabilities.

Completeness

It explicitly answers both what ('Audit experiment integrity... to check for fake ground truth, score normalization fraud, phantom results, and insufficient scope') and when ('Use when user says... or after experiments complete before writing claims') with concrete trigger phrases, matching the anchor-5 example exactly. Third person voice is used throughout.

Trigger Term Quality

Natural trigger phrases are present in two languages ('audit results', 'check experiment integrity', '审计实验', '实验诚实度') plus a temporal trigger ('after experiments complete before writing claims'), but common English variations such as 'validate results' or 'verify results' are missing. Good keyword coverage with a few natural terms absent fits anchor 4 rather than 5's comprehensive synonym coverage.

Distinctiveness Conflict Risk

A clear niche — experiment integrity fraud auditing via cross-model review — with distinct, domain-specific triggers ('fake ground truth', 'score normalization fraud', 'phantom results') unlikely to fire for other skills. Minimal conflict risk.