Reading the response

Every answer comes back under the name you gave its question, as numbers you can act on directly. What to do with them is in Acting on confidence.

{
  "model": "hartley-1",
  "answers": {
    "followed_policy": {"type": "noul", "noul": 0.04, "abstain": 0.01},
    "next_step": {
      "type": "choice", "choice": "escalate",
      "probabilities": {"escalate": 0.88, "close": 0.03, "ask_user": 0.09},
      "confidence": 0.82, "abstain": 0.02
    }
  },
  "usage": {"input_tokens": 142, "output_tokens": 0},
  "latency_ms": 41.7
}

Field	Meaning
`answers.<name>`	One answer per question, under the name you gave it.
`noul`	For yes/no questions: the probability of yes.
`choice` / `score`	The most likely option, or the expected level (0-based).
`probabilities`	The calibrated distribution over your options (abstain excluded, renormalised). An answer given at 0.9 is right about 90% of the time.
`confidence`	How peaked the distribution is: `(K·p_max − 1)/(K − 1)`. 1 = all mass on one option, 0 = uniform.
`present`, `value`, `span`	For `extract`: whether the value is stated, the text the model pointed at, and its `[start, end)` character offsets. For a string state they index into your string; for a JSON state they index into Hartley's text rendering of it, so use `value` rather than slicing.
`candidates`	For `extract`: the most likely spans with their probabilities, best first.
`abstain`	The "can't tell" mass: how much the model thinks the state lacks the evidence. Route high-abstain answers to a person or a larger model.
`usage.input_tokens`	What you are billed for. Output is free.
`latency_ms`	How long the model took, in milliseconds.
`calibration`	Which calibration was applied: `general`, or `none` when you asked for raw probabilities.

Every response carries an x-request-id header. Include it when you contact support.