Guardrails that stop things.

PII Shield and content policy run before the upstream call is made — a blocked request returns a clean 400 and is never billed.

Enforced inline, not logged after the fact: the model never sees what the policy rejected.
Observe mode first — see what would have blocked against real traffic before enforcement goes on.
A preset template library plus custom patterns, with per-match suppression and export.

Stopped, not noted.

A card number goes into a prompt. The request never reaches the model, comes back as a clean 400, and costs nothing — because the check runs before the upstream call, not in a report you read on Thursday.

See it before you enforce it.

Turn a policy on in observe mode and watch what it would have blocked against real traffic. When the numbers look right, promote it. Nobody has to guess how much a new rule is about to break.