Topics · Domain workflow
Auditors verify which dataset commitments and pipeline claims were signed, whether evaluation artifacts match roots, and where gaps exist in the staged trail.
Concrete scenario
A deployed classifier is flagged for bias after six months in production. Compliance asks for proof of the training corpus version and every transformation before deployment. The ML platform was migrated twice; only partial run logs and an unverified export survived.
Problem
Model audits fail slowly. When harm appears in outputs, teams must reconstruct which dataset version, filtering steps, and evaluation artifacts actually formed the training path — not which slide deck said they did.
Mechanism
Training and audit workflows emit receipts at dataset snapshot, transform, train, and evaluate stages. Each stage binds manifests and evaluation roots so later reviewers can traverse the path without the original MLOps UI or vendor retention policy.
Verifiable outcome
Scope boundary
Receipt trails prove signed pipeline facts — not model fairness, regulatory clearance, or that every GPU hour was receipt-instrumented unless operators enforce that profile.