Safety Models And Governance

An agent pins a specific GroqCloud model id; the model is deprecated and removed, and every request begins returning model_not_found.

Track model lifecycle: monitor deprecation notices, resolve model ids from GET /models rather than hardcoding, and have a migration path to a supported successor. Re-baseline evals against the new model before cutting over. Do not assume a pinned model id is permanent.

Llama Guard flags a prompt as unsafe but the agent logs the verdict and proceeds with the completion anyway.

Enforce the moderation verdict: when content is flagged, block or route to a safe response per policy rather than proceeding. Map the model's safe/unsafe verdict and category to a concrete action; a logged-but-ignored verdict provides no protection.

An evaluation suite passes today but silently regresses after Groq updates the underlying model behind the same id, with no version pinning or fingerprint check.

Pin the exact model id, capture any system_fingerprint/version signal, and re-run the eval suite on model updates to detect behavior drift before it reaches production. Treat a model update as a change that requires re-validation, not a no-op.