OtherAdvancedClaude API
Authorize actions with side effects using deterministic rules
A model can classify a dubious input, but allowing a refund or a shipment has to be a decision you can test and reproduce.
Place the control according to the point: input screening can use a lightweight classifier for jailbreaks, output screening a judge for toxicity or policy, and tool authorization should almost always be deterministic (allowlist, identity, scope). No single control covers everything: chain them and know what each one lets through.
Example
Before running issue_refund: - Does the user have permission for this action? - Is the amount under the cap? - Is the action on the allowlist? If any fails, it does not run, whatever the model says.