Lean operational prose that assumes Claude's competence — no basic-concept padding about what models or metrics are — and every section (tolerance direction, segment behavior, limitations) earns its place for a genuinely multi-step gate.

Each step ships fully executable, copy-paste-ready Python with concrete imports and a complete GitHub Actions YAML; the regression-vs-classification variant is shown inline rather than left as pseudocode.

Steps 1–8 are clearly sequenced with explicit validation checkpoints (per-metric FAIL, segment gate FAILED, train-test degradation gate FAILED), each printing a result and calling sys.exit(1); the anti-patterns and limitations tables provide error-recovery guidance.

SKILL.md is a runnable overview; the only externalized detail — YAML threshold config and full metric signatures — lives in references/gate-config-and-metrics.md, a real one-level-deep file signaled with explicit links in Steps 3 and 8.

Lists multiple concrete actions ('Computes held-out metrics (accuracy, F1, AUC, RMSE)', 'compares them against the current production model', 'failing promotion when any metric regresses beyond a configured tolerance', 'Adds per-segment checks via Deepchecks WeakSegmentsPerformance').

Clearly answers both 'what' (compute and compare held-out metrics, per-segment checks, block promotion) and 'when' via an explicit 'Use when a retrained model is a candidate for promotion and the CI pipeline must enforce a per-metric pass/fail gate' clause.

Uses natural terms an ML engineer would say ('retrained model', 'production model', 'promotion', 'CI pipeline', 'per-metric pass/fail gate', 'model registry'), reinforced by the keyword metadata, with no generic jargon-only phrasing.

Scoped to a narrow niche — candidate vs. production pairwise comparison with per-metric tolerances and segment gating — with distinct triggers unlikely to fire for adjacent skills like deepchecks-tests or evidently-monitoring.