Content filtering checks model inputs or outputs for disallowed material. It is one layer of defense in depth; it does not replace policies that control which actions an assistant may take. section 6.6 section 10.7
In an assistant workflow, filtering can happen at multiple points:
- Check incoming messages before they reach the model; a topic classifier can block requests such as medical advice. section 10.7
- Filter the model’s response before returning it; output checks can catch hallucinations or redact personally identifiable information. section 6.6
Figure 6.9 Execution policy guardrails control what the system does, and content guardrails filter what the model says. Both are necessary, but only execution policies can enforce business-specific rules.

Contrasts filtering what a model says with execution policies that control what the system does. section 6.6




