Guardrails constrain model inputs and outputs to keep behavior safe and compliant - blocking disallowed content, redacting sensitive data, validating formats, and refusing unsafe requests. They sit around the model, not inside it, and should be tested like any other control.
Why it matters
Models will process whatever they are given. Guardrails are the policy layer that keeps inputs and outputs within acceptable bounds.
How it works
Checks run around the model - validating and redacting inputs, filtering or flagging outputs, and refusing disallowed requests. They are ordinary controls that should be tested and versioned like any other code.