Imported from NeerajCodz/ml-stack (adapters/omp-marketplace/plugins/ml-stack/skills/ml-stack-audit/SKILL.md). Install upstream with npx skills add NeerajCodz/ml-stack --skill ml-stack-audit. Copyright stays with the author.

Independent release audit

Require requirements lock, data audit, validation report, research claims, model revisions, run and tracking records, artifact hashes, privacy/licensing evidence, deliverable schema, promotion state, and immutable manifests.

Workflow

1. Build an evidence index with exact paths, URLs, revisions, hashes, timestamps, and owners.
2. Reconcile every requirement, claim, dataset split, model/tokenizer, environment, metric, artifact, and approval against measured evidence.
3. Check reproducibility, leakage, privacy, license, schema, provenance, and promotion gates; label measured, derived, unknown, and contradicted values.
4. Review release target and rollback state independently; do not accept the experimenter's or deployer's self-approval.
5. Return exactly PASS, FAIL, or ABSTAIN, blocking evidence, remediation, and next owner: ml-stack-deployment only after PASS, otherwise the named upstream owner.

Boundary and outputs

Never fill missing evidence with assumptions, mutate artifacts, or approve an unobserved operation. ABSTAIN means evidence is insufficient; FAIL means a known gate is violated. Return an audit report with exact paths and hashes. Next owner is ml-stack-deployment after PASS, ml-stack-experiment for reproducibility gaps, ml-stack-data for leakage/privacy gaps, or ml-stack-hub for metadata/licensing gaps.

See audit report, claims and provenance, release gates, and privacy and licensing.