Guardrail
In shortAn action taken on the device as AI is used: record, warn, redact or block.
Where it sits in EndLayer · step 6 of 7, decision
What it means
Guardrails run in the Layer 1 Agent, so they act before a prompt leaves the machine. Monitor records quietly, warn tells the person, redact removes only the sensitive part, and block stops it with a stated reason.
Each one becomes an event with the policy version behind it.
Why it matters
Some risks need stopping as they happen, not reporting a week later.