Model Registry
Advanced 16 min read Model Registry

Governance and Auditing

Make every model decision defensible

Governance is who may do what, with what evidence; auditing is proving it happened. Registry rules, permissions and audit logs together make model decisions defensible.

model governanceauditapprovalspermissionsevidencecompliance
Prerequisites: Registration and Stage Rules Model Metadata and Lineage
Before we start

What you will learn

  • Define governance roles and permissions in the registry
  • Enforce evidence requirements and approvals
  • Build audit reports regulators can rely on
Introduction

What is this?

Model governance is the policy framework — roles, permissions, evidence requirements, approval flows — applied to the registry so model changes are deliberate. Auditing is the record that proves those policies were followed: every registration, transition, alias flip and deployment logged with actor, timestamp and evidence.

Motivation

Why do you need it?

  • Regulated industries (finance, health) must answer 'who decided this model serves and why.'
  • Incidents resolve faster when the change record exists and is queryable.
  • Permissions prevent the 'everyone can promote' chaos that breaks production trust.
  • Consistent evidence requirements raise the quality floor of shipped models.
Think about it

Mental model

Imagine

A bank's dual-approval rule for large transfers

Think of

Model governance controls

One clerk can't move millions alone — a second approver and a full ledger entry are required. Models in production are similarly consequential; promotion needs approvals, and the ledger (audit log) is non-negotiable.

Mechanics

How it works

  • Roles are defined: data scientist (register), reviewer (approve), platform admin (manage), read-only (audit).
  • Permissions map to registry operations per role and stage.
  • Evidence requirements are attached to transitions (metrics, split, shadow agreement).
  • Every registry mutation writes an audit event: actor, action, version, timestamp, evidence refs.
  • Audit reports enumerate live models, who promoted them, and with what evidence.

Governance controls over the registry and its audit trail.

Roles & permissions
Evidence rules
Operations
Audit log
Reports
Real world

In production

  • A bank requires two approvers for Production and archives the evidence with the transition.
  • Quarterly audit output: every serving model, its version, owner, promotion date and evidence URLs.
  • Access review: data scientists can register and stage, only designated reviewers promote to Production.
Guidelines

Best practices

  • Least privilege: roles map to the narrowest operations needed.
  • Require evidence on promotion, and validate it automatically where possible.
  • Append-only audit logs; write once, never edit or delete.
  • Review access quarterly and reconcile registry state against running services.
Watch out

Common mistakes

  • Permissions so locked down that bypasses become unavoidable.
  • Evidence fields optional, so promotions happen with empty rationale.
  • Audit logs in a place that can be edited or truncated.
  • Governance scoped to the registry but serving/monitoring changes ungoverned.
When things break

Debugging tips

  • 'Who promoted that model?' — query the audit log for the transition; don't ask around.
  • Vulnerable change detected — the evidence reference should point to the exact metrics and run.
  • Approval stuck — verify the required approver role and that evidence fields are all populated.
Interviews

Interview checkpoint

Test yourself

Quick quiz

0/2
1

A regulator asks who approved the current production model and on what evidence. Where do you point?

2

Audit logs that can be edited by ordinary users are acceptable.

Go deeper

Official documentation

MLflow registry permissions docs
docs
NIST AI Risk Management Framework
docs