Confidence-Gated Action Router

{
  "model": "jev-latest",
  "state": {
    "user_message": "{{user-message}}",
    "account_context": "{{account-context}}"
  },
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": "What action is the user asking the assistant to perform",
      "criteria": {
        "{{read-only-action}}": "A read-only lookup that changes nothing",
        "{{risky-action}}": "An action that moves money, deletes data, or cannot be undone",
        "other": "Something else, unclear, or small talk"
      }
    },
    "confirmed_intent": {
      "type": "noul",
      "instructions": "The user states the action plainly rather than asking whether it is possible or thinking out loud"
    }
  }
}

Decision rules to implement in code, not in the model:
1. intent.confidence below 0.6: hand off to a human or ask one clarifying question. Never act.
2. intent.choice is the read-only action and confidence is 0.6 or above: execute immediately.
3. intent.choice is the risky action and confidence is between 0.6 and 0.85: show the user exactly what will happen and ask them to confirm.
4. intent.choice is the risky action, confidence above 0.85 and confirmed_intent above 0.8: execute, then log the two scores next to the action for audit.

Raise every threshold before lowering any. Tune them on a sample of your own real messages, starting strict and loosening only where the false-positive cost is low.