Publish and validate a model

Submit a model for publication, and what Model Risk Management needs to validate it at its risk tier.

For: Builder / Publisher, Model Risk & Governance · 1 min read

What to prepare

A model card: intended use, out-of-scope use, limitations, training data and lineage.
Your development and test results, and the pipeline that produced them.
For credit models: fairness results on protected groups and reason codes for every output.
For models replacing an existing one: parallel-run results against the champion.

1Register the model in the Vertex AI Model Registry from your pipeline.
2Open Publish, choose Model, and fill the metadata, classification and proposed risk tier.
3Complete the responsible AI checklist.
4Submit. The Assurance Agent re-runs your tests and checks the pack is complete; an incomplete pack is returned with the failing field named.
5The Validation Case Agent prepares the case for Model Risk Management. A named validator runs the validation for the tier and signs the opinion.
6Once approved, MLOps deploys the validated version. It appears in the catalogue with its validation status.

It depends on the tier and the completeness of the pack. See the FAQ for typical durations.