Controls

Policy Checks

Run checks before an agent speaks, writes to a system, escalates a case, or calls an expensive model for a second pass.

POST /v1/checks
{
  "workflow": "support_escalation",
  "proposed_action": "send_customer_reply",
  "risk": {
    "topic": "refund_policy",
    "confidence": 0.71
  },
  "transcript_ref": "conv_2841"
}

Policy checks should be narrow enough to make a concrete routing decision and broad enough to avoid scattering review logic across agent prompts.