Conciseness

The 87-line body is dense throughout: a commented command block, mode-selection policy, gate conditions, bolded hard rules with one-clause rationales, a script table, and a references list. The only prose that is not instruction is the four-line intro, and each of its sentences justifies a component (repeated trials, path-vs-outcome grading), so no section could be cut without losing a rule's motivation. This matches anchor 5 — lean, assumes Claude's competence — rather than 4, which would require over-explanation that could be trimmed.

Actionability

All three workflow commands are complete and copy-paste ready with real flags (--spec, --k 3, --quarantine-out, --baseline, --candidate, --quarantine), the required JSONL row schema is shown inline, and the scripts table documents exit codes per script; the claimed "All support --sample and --output json" was verified against the actual scripts. This matches anchor 5 — fully executable, common cases covered (sample data files exist in assets) — not 4, since even step 0's output format is specified exactly.

Workflow Clarity

Steps 0-3 are numbered, sequenced, and paired with commands; the five explicit "The gate blocks on" conditions plus the warnings for underpowered drops and cost rises give the skill explicit validation checkpoints with machine-enforced exit codes. Although this is a batch-operation skill, the feedback-loop cap does not apply: recovery paths are specified (quarantine entries with owner/expiry re-enter the gate automatically, the power note says how many cases are needed when the sample is too small). Matches anchor 5, not 4, because checkpoints are explicit and enforced rather than merely listed.

Progressive Disclosure

SKILL.md is a true overview: workflow, gate conditions, and hard rules inline, with deep material split into three reference files each introduced by a one-line content summary (e.g. "pass@k vs pass^k, per-case averaging, paired CI, McNemar, power (7 sources)"). All seven referenced bundle files were verified to exist, and the references are one level deep (they cite script names, not further documents). Matches anchor 5 — clear overview, well-signaled one-level-deep references, appropriate split — not 4, since every reference is signaled with what it contains and when to open it.

Specificity

"Runs repeated trials, gates candidate vs baseline with paired statistics in regression_gate.py, checks tool-call trajectories and finds flaky cases" lists four concrete actions that map one-to-one onto the skill's components (trial running, statistical gate, trajectory grading, flake detection), even naming the gating script. This matches the anchor 'lists multiple specific concrete actions; comprehensive coverage' rather than 4, since there are no coverage gaps — every capability in the bundle is named.

Completeness

The 'what' is explicit ("runs repeated trials, gates candidate vs baseline with paired statistics..., checks tool-call trajectories and finds flaky cases") and the 'when' is an explicit Use clause with concrete triggers ("Use before shipping any agent, prompt, model or tool change, or when a pass-rate drop needs a verdict"). This clearly matches anchor 5 — both questions answered with concrete trigger phrases — not 4, where the 'when' would be present but less specific.

Trigger Term Quality

Natural phrases are present — "before shipping any agent, prompt, model or tool change", "pass-rate drop", "finds flaky cases" — which users would plausibly say. It falls short of anchor 5 because common synonyms like "regression", "eval", "evaluation", or "benchmark" are absent, so a user asking about a 'regression' or 'agent evals' in natural language would not hit these words; it is above anchor 3 because the terms present are varied and natural rather than domain-generic.

Distinctiveness Conflict Risk

The functional core ("gates candidate vs baseline with paired statistics", "checks tool-call trajectories", "pass-rate drop needs a verdict") carves a clear niche distinct from generic eval or analysis skills. It sits at anchor 4 rather than 5 because the trigger "before shipping any agent, prompt, model or tool change" is broad enough to overlap with adjacent prompt-governance or CI/deploy skills; it is well above anchor 3 since the statistical-gating actions are highly distinctive.