What is the minimum security content a 'model card' should include to support AI risk management for a deployed production model?

• A.A model card must include the model's full architecture diagram and weight sizes. Canary deployments catch adversarial behaviour automatically, since a poisoned model diverges from baseline on the first mirrored request.
• B.The model's accuracy on its primary benchmark and the date it was trained.
• C.Model cards are optional documentation; the only required artifact is a performance benchmark score.
• D.Intended use cases and known out-of-scope uses; performance metrics across demographic subgroups; training data sources and known data limitations; evaluation methodology; known limitations, failure modes, and biases; recommended deployment constraints; and version/update history.