model governance & audit

When a model decides who gets a loan, a diagnosis, or a parole recommendation, the organization must be able to prove, after the fact and to an outside party, that the decision was made responsibly. Model governance and audit is the set of controls, approvals, and records that makes a model accountable: who built it, on what data, with what known limitations, who approved it for which use, and an immutable trail of every version and decision so the whole lifecycle can withstand scrutiny.

Concretely, governance adds gated approval workflows (a model cannot reach production without sign-off from defined roles), structured documentation such as model cards and datasheets that state intended use, performance across subgroups, and known risks, and tamper-evident audit logs binding each prediction to the exact model, data, and code that produced it. It enforces access controls and segregation of duties, schedules periodic reviews and bias and fairness assessments, and maps all of this to regulatory regimes such as the EU AI Act, sectoral model-risk rules, or internal policy.

The reason it is hard is that accountability must be designed in, not bolted on. Lineage, reproducibility, and immutable records have to exist before the audit is requested, because you cannot retroactively prove how a prediction was made if you did not capture it at the time. Good governance leans on the same versioning, metadata, and registry infrastructure as the rest of MLOps, repurposed for evidence rather than convenience.