Which are essential components of AI model governance in CPMAI?

• A. Model registry and versioning only.
• B. Data privacy impact assessment and regulatory compliance.
• C. Version management of code and license tracking.
• D. Model registry, versioning, lineage, performance monitoring, bias audits, and retirement criteria.

Answer: (D)

In CPMAI, AI model governance hinges on overseeing the full lifecycle of a model to ensure it remains reliable, fair, and compliant from creation to retirement. The best way to achieve this is to include a comprehensive set of components that cover discovery, change control, traceability, ongoing performance, fairness, and decommissioning.

A model registry serves as a central catalog of models and their metadata, enabling reproducibility and easy discovery. Versioning keeps track of every change to both the model and its code, so you can pinpoint when updates were made and why. Lineage records trace how data flows into the model, what features are used, and how inputs are transformed, which is crucial for audits, debugging, and regulatory reviews. Ongoing performance monitoring watches how the model behaves in production, detects drift or degradation, and signals when interventions are needed. Bias audits assess fairness and potential disparate impacts, helping to prevent harmful outcomes and satisfy ethical and regulatory requirements. Retirement criteria define when a model should be decommissioned or replaced, ensuring that outdated or unsafe models do not continue to run.

This combination is essential because it provides end-to-end governance—tracking what exists, how it changed, how it operates in the real world, and when it should be retired—supporting accountability, risk management, and compliance throughout the model’s life.

Elements like a narrow focus on registry and versioning alone miss lineage, monitoring, fairness assessments, and retirement planning. Data privacy impact assessments and regulatory compliance are important but address only specific risk areas and ongoing governance needs, not the full lifecycle. Version management of code and license tracking covers changes and licenses but not the model artifacts, data lineage, performance in production, or retirement decisions.