The confidence threshold for triggering guardrail actions based on Automated Reasoning policy violations.