Assumes Claude's competence throughout - it states the Fairlearn/EEOC/GDPR conditions by quoting and linking the source rather than re-explaining what fairness or a PDF-style concept is, and each section adds only what Claude would not already know (the matrix, the bands, the rules).

Provides copy-paste-ready, executable specifics: the verdict-band table (DPD <= 0.05 -> PASS, etc.), the four-field waiver schema ("Reason:, Approved-by:, Re-review-date:, expires:"), the six evidence rules R1-R6 with explicit triggers, and a worked example emitting a concrete coverage report.

An explicit eight-step sequence (assign tier -> read required evidence -> name metric -> accept conflict -> apply bands -> require intersectional counts -> triage scan -> check rules) with validation checkpoints built into the rules, e.g. R3 (BLOCK band names a missing artifact) and R4 (a monitoring claim without a reference dataset is contradicted).

SKILL.md is a self-contained overview that signals exactly one one-level-deep reference - [references/worked-example.md](references/worked-example.md) - which exists as a real bundle file, keeping the dense end-to-end walk-through out of the main body while the matrix and rules stay inline.

Lists multiple concrete actions - "Assigns a machine learning model to a low, medium, or high risk tier", "derives the fairness and explainability evidence", "group metrics per declared sensitive feature, intersectional breakdowns with per-cell counts, vulnerability scan categories, a drift monitoring plan, and per-prediction explanation logs" - each naming a distinct artifact rather than vague capability.

Answers both halves explicitly: the what ("Assigns a machine learning model to a... risk tier... then derives the fairness and explainability evidence that tier must produce") and the when via an explicit "Use when a model release candidate is up for promotion... or when a model card declares a risk tier and the attached evidence bundle has to be checked".

Covers natural terms a reviewer would say - "model release candidate is up for promotion", "model card declares a risk tier", "evidence bundle", "fairness artifacts" - matching how practitioners phrase the request.

Occupies a clear niche (risk-tier-to-evidence-requirement mapping) with triggers tied to release promotion and model-card tier declarations, unlikely to fire for general fairness or model-training skills.