Guardrails

A guardrail checks text going to or from an AI system: what people send to it, what the model answers, or both. When it finds what it looks for, it fires, and does one of three things. Every AI system has the same eight guardrails, all off until you switch them on. See Guardrail types.

​

Draft and live version

You change an AI system’s guardrails in its draft. Save draft keeps your changes without enforcing them. Apply turns the saved draft into the next numbered version and makes it the live version. A version never changes once applied. To go back, roll back to an earlier version, which applies it again as a new version. See Versions. The AI system’s Guardrails tab always shows what is live: a badge reads Live: version N or Nothing is live, and Changes saved, not applied when the draft differs from the live version. You can always tell what is enforced now, what is only drafted, and what is waiting for you.

​

Where guardrails come from

• From an assessment. Submitting an assessment switches on the guardrails its controls map to, in Log only and with nothing chosen yet. The assistant then suggests settings from your answers. See Guardrail suggestions.
• By hand. Owners, admins and the AI system’s owner can configure guardrails at any time, with or without an assessment.

​

Who can change them

Every member can see an AI system’s guardrails and versions. Owners, admins and the AI system’s owner can save, test, apply and roll back. Anyone else sees “Read only. Owners, admins and the owner of this AI system can change these.”

​

Whose guardrails

An AI system has its own guardrails, and so does your organization. The organization’s apply to traffic that does not belong to a registered AI system. See Organization guardrails.

​

Where they are enforced

The live version is what the AI Gateway enforces on your applications’ requests: applying a version sends it there. See Guardrails at runtime. Use Test guardrails to check what a draft would do before you apply it.