1. Contoso Financial is building a validation framework for a custom Azure Machine Learning regression model that predicts credit-limit recommendations surfaced through a Dynamics 365 Sales Copilot panel. The model will be retrained monthly on new transaction data. The Responsible AI review board requires that validation criteria be formally defined before the model is promoted to production after each retraining cycle. Which TWO criteria should the architect include as mandatory gates in the model validation process? (Select TWO.)
- A. The retrained model must achieve equal-or-better performance on a held-out fairness dataset segmented by protected demographic groups compared with the currently deployed model version.✓ Correct
- B. The model's Docker container image must pass a vulnerability scan with zero critical CVEs before the model endpoint is updated in Azure Machine Learning.
- C. A human-in-the-loop review must be completed for every individual prediction the model makes before the result is displayed in Dynamics 365 Sales.
- D. The retrained model must meet or exceed baseline regression metrics (e.g., RMSE and MAE) on the designated test dataset before promotion.✓ Correct
- E. The model training run must complete within a fixed compute-cost budget defined in the Azure Machine Learning compute cluster configuration.
- F. All training data must be re-labelled by a human reviewer after each monthly ingestion to eliminate potential label drift before retraining begins.
Explanation
Fairness evaluation against protected demographic groups (A) is a mandatory Responsible AI gate that ensures the model does not produce biased credit recommendations — a regulatory requirement in financial services. Baseline regression metric thresholds (D) such as RMSE and MAE are standard quantitative quality gates that confirm the retrained model has not regressed in predictive accuracy before promotion. Together these address both performance and fairness. — Container vulnerability scanning (B) is an important DevSecOps practice but is an infrastructure gate, not a model validation criterion; it does not validate model quality or fairness. — Human-in-the-loop review for every prediction (C) is operationally infeasible at scale and is not a validation criterion for the model itself; human oversight is better applied at high-risk decision thresholds. — Compute cost budget compliance (E) is a cost-governance concern, not a model quality or responsible AI validation gate. — Requiring full human re-labelling of all data monthly (F) is disproportionate, not a standard validation criterion, and would make the monthly retraining cycle impractical.