Question

A financial services firm uses a custom-tuned model in Microsoft Foundry to score loan applications. Regulators require that any change to the model's training data or hyperparameters be attributable to a named individual, reviewed before promotion, and reconstructable months later. The data science team currently retrains ad hoc and promotes directly to production. As the solution architect, which change to the ALM process best satisfies the regulatory requirements without blocking legitimate iteration?

Accepted Answer

Introduce a gated deployment pipeline that versions training datasets and model artifacts, records the identity and approver of each change, and retains immutable audit logs of dataset and hyperparameter modifications.

Answer

Enable Foundry content safety filters on the model endpoint and log all inference requests to Application Insights.

Answer

Restrict model retraining to a single senior data scientist and require them to email a summary of each change to compliance.

Answer

Pin the production model to a fixed version and disable retraining entirely until the annual compliance review.