Guardrails

Guardrails are policies and filters around an LLM — input screening, output validation, topic allowlists, PII redaction, and tool-permission checks.

They complement (not replace) model alignment. Defense in depth: validate schemas, constrain tools, and monitor production traces.