Guardrail policies are rules, written in plain language, about what an AI system may write or do. Each one is paired with the action to take when it is broken, such as block, redact or flag.
What is in a policy
- The rule. A sentence that can be answered yes or no about one message, reply or tool call.
- The action. What happens on a yes. See guardrail actions.
- The scope. Where it applies: inputs, outputs or tool calls, and for which feature or users.
The verdict comes from a checker, often a guard model.
How to write a good one
Name the AI's own act. A guardrail sees what the AI is sent, what it writes and what its agent does, so the rule has to be about one of those.
Keep one rule per policy: a bundle is hard to tune, and a vague rule produces false positives. Run a new policy over past traffic before it goes live (policy backtesting).
With White Circle
Policies in Control are written in plain language and enforced on every input, output and tool call. They are unlimited, built-in or fully custom, and versioned: an update applies at once and a rollback is one click.
