Conciseness

The body is ~75 dense lines with no padding: no concept tutorials, no re-explanation of statistics basics, and judgment stated in single sharp lines like "Dropping a case is the cheapest fake fix" and "A user meets every trial, so pass^k is the reliability number". The pass@k/pass^k paragraph defines the skill's own metric convention rather than teaching known material, so every token earns its place — the level-5 anchor; level 4 would require identifiable trimmable over-explanation, and none stands out.

Actionability

The workflow block gives three copy-paste-ready commands whose flags (--spec, --results, --k, --quarantine-out, --baseline, --candidate, --quarantine) verifiably match the actual argparse interfaces of the bundled scripts, plus a concrete JSONL row schema ("{\"case_id\": ..., \"trial\": ..., \"passed\": ..., \"trajectory\": ...}") and a per-script exit-code table. This is fully executable guidance covering the common cases — the level-5 anchor; there are no pseudocode or missing-detail gaps that would justify level 4.

Workflow Clarity

Steps 0–3 are clearly sequenced (run trials → grade trajectories → detect flakes → gate), validation is explicit via documented exit codes ("Exit 4 blocks the merge"), and the "gate blocks on" list is a five-condition checklist. Feedback loops for error recovery are present — "An expired entry re-enters the gate automatically", "the gate says how many cases it would take" for underpowered drops, and "It warns, and a human decides" — matching the level-5 anchor (explicit validation, feedback loops, checklists).

Progressive Disclosure

The body is a genuine overview: workflow, rules, a script table, and a references section where every path (3 references/, 3 scripts/, 4 assets/) verifiably exists, is one level deep, and is annotated with its content (e.g. "pass@k vs pass^k, per-case averaging, paired CI, McNemar, power (7 sources)"). Statistics detail is appropriately pushed to the reference files rather than inlined — the level-5 anchor; the level-4 anchor's "minor organization gaps" are absent.

Specificity

The description lists four concrete actions — "Runs repeated trials", "gates candidate vs baseline with paired statistics in regression_gate.py", "checks tool-call trajectories", "finds flaky cases" — each naming its object and even the implementing script. This matches the level-5 anchor (multiple specific concrete actions, comprehensive for the domain); the level-4 anchor ("minor gaps in coverage") does not apply because trial-running, statistical gating, trajectory grading, and flake handling together cover the skill's full scope.

Completeness

Both questions are answered explicitly: the "what" is "Runs repeated trials, gates candidate vs baseline with paired statistics... checks tool-call trajectories and finds flaky cases", and the "when" is an explicit trigger clause — "Use before shipping any agent, prompt, model or tool change, or when a pass-rate drop needs a verdict". This mirrors the level-5 anchor example (concrete what plus "Use when..." with concrete trigger phrases); level 4 would require the when-clause to be only partly explicit, which it is not.

Trigger Term Quality

Natural phrases a user would actually say are present — "before shipping", "agent, prompt, model or tool change", "pass-rate drop", "flaky" — giving good keyword coverage. It falls short of the level-5 anchor (comprehensive synonyms and variations) because the most common user vocabulary for this domain — "eval(s)", "evaluation", "regression", "benchmark", "A/B" — never appears in the description text, so a user asking "did my evals regress?" must paraphrase to match.

Distinctiveness Conflict Risk

The capability framing (repeated trials, paired statistics, regression gating) carves out a clear niche, but the opening trigger "Use before shipping any agent, prompt, model or tool change" is broad and overlaps the invocation space of closely related skills (prompt governance, single-session self-eval). That is precisely the level-4 anchor ("mostly distinct; minor overlap risk with closely related skills"); level 5 would need the distinct trigger to carry the whole clause, as the PDF example's file-type triggers do.