Versioning and Governance Tier

• Encode promotion workflows in the registry
• Show how lineage links datasets, runs, code and models
• Design audit trails that satisfy compliance review

• Compliance (SOC 2, HIPAA, GDPR, model-risk review) requires demonstrating controlled, reviewed model changes.
• Rollback needs versioning: you can't revert to what you didn't store.
• Automation needs gates: CI/CD can auto-promote only because the registry encodes what 'ready for production' means.
• Lineage answers the incident question 'what trained this?' without a hunt across notebooks.

Every flight is tied to a maintained, signed, logged airframe; you fly because the log proves it. Production models fly the same way.

• Training writes a versioned artifact plus metadata (metrics, data version, code commit, framework, signature) to the registry.
• A promotion workflow moves the model through stages — staging → production — with required approvals and quality gates at each hop.
• Every promotion, deployment and rollback is recorded as an immutable event with actor, time and artifact version.
• Deployments reference a registry version and stage, so the serving tier and the audit log agree on what is live.
• Governance reports — model inventory, lineage trees, promotion history — are generated from the same records.

• A bank's model-risk review pulls a lineage report per model before any production promotion — the entire package is generated from registry records, not spreadsheets.
• An e-commerce CI pipeline auto-promotes a model to production only when offline metrics pass and the data version is current; humans approve only the final rollout.
• A medical-device vendor keeps every deployed version for 7 years; rollback is a one-command deploy of a registry-backed artifact.

• Make the registry the single source of truth for what is in production; serving pulls by version, never by ad-hoc file path.
• Encode, don't narrate, the promotion policy: quality gates run, approvals are recorded, non-authorized paths fail.
• Attach data version and code commit to every model so lineage is complete and rebuildable.
• Immutable artifact store: never overwrite a version — rollback needs the old bits to still exist.

• Versioning the model file but not its dataset or code — a 'versioned' model that can't be reproduced is a versioned lie.
• Human approval as an email thread outside the registry, so the audit log and reality disagree.
• Letting teams deploy arbitrary files alongside the registry, quietly bypassing governance.

• In an incident, pull the lineage for the live version first — it names dataset, code and metrics in one query.
• If a rollback fails, check that the previous artifact wasn't garbage-collected.
• Diff the audit log against actual deployments regularly; divergence means a bypass path exists.