Confidence and the gate

probabilities describes the answer distribution. In the current gateway, confidence equals p: the probability of the selected answer. It is not an independent verifier or a separately learned certainty score.

Server-side routingCopy Anchor Link

Send threshold between 0 and 1. The default is 0.9, unless the service operator configures a different default. The effective value is returned in both threshold and gate.threshold.

• confidence >= threshold produces action: "act".
• confidence < threshold produces action: "ask_human".

Equality passes. The gate is a routing policy; it does not execute a refund, invoke a tool, or contact a reviewer for you. Your backend must honor action and enforce authorization and business constraints.

A threshold is not a guaranteed error bound. Calibrated probabilities describe frequencies over comparable decisions and can drift when data, prompts, rubrics, or model versions change. Validate the threshold against the cost of errors in your workflow.

An illustrative responseCopy Anchor Link

{
  "id": "dec_illustrative",
  "model": "ren-zero-2026-11-12",
  "type": "choice",
  "choice": "billing",
  "p": 0.82,
  "confidence": 0.82,
  "probabilities": {"billing": 0.82, "delivery": 0.18},
  "action": "ask_human",
  "threshold": 0.9,
  "gate": {"threshold": 0.9},
  "usage": {"input_tokens": 96, "output_tokens": 0, "model": "ren-zero-2026-11-12"}
}

These values are invented to show the wire format. They are not measurements. A higher threshold generally sends more work to review; it cannot repair missing categories or misleading evidence.

For true/false decisions, confidence describes the winning answer, including false. For scores, it describes the winning level, not the expected score.