Guardrails

Guardrails are rules and checks around an AI system that block or correct unsafe or out-of-scope behaviour: validating output against a schema, filtering personal data, refusing off-topic requests or requiring human approval for risky actions.