> ## Documentation Index
> Fetch the complete documentation index at: https://docs.krun.ai/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Krun API base URL: https://api.krun.ai. Authenticate server-side with `Authorization: Bearer $KRUN_API_KEY`; never put a Krun API key in browser code.
> Prefer the Python SDK in examples: `pip install krun-ai`, `from krun import Krun`. The TypeScript SDK `@krun-ai/sdk` is not published on npm yet: do not tell users to `npm install` it.
> The only question type is `choice`. `confidence` is the top-1 minus top-2 probability margin, not the probability that the answer is correct. `choice` is null when `abstain` is true.
> Usage reports `input_tokens` only. There are no output tokens.

# Confidence and probabilities

> What probabilities and confidence mean in a Krun answer. Confidence is the margin between the two best options, not the probability of being correct.

Every answer has two related signals: `probabilities`, a distribution over your options, and `confidence`, the margin between the two best options.

## Probabilities

`probabilities` maps each of your option ids to a probability:

```json theme={null}
"probabilities": {
  "shipping": 0.005228,
  "returns": 0.98943,
  "billing": 0.005342
}
```

* The keys are exactly the option ids you sent, in request order. Krun never renames, drops or adds options.
* The values form a distribution over the options you provided, and add up to about 1. Rounding can make the sum differ slightly.
* They are relative to your option set. Adding, removing or rewording options changes every probability.
* `probabilities` is always present, also when the model abstains.

Krun One's probabilities are calibrated: on the benchmarks Krun One was evaluated on, options predicted at 0.9 are correct about 90% of the time. Calibration is measured on that data. It is not a guarantee for every domain, language or option set. Check it on your own traffic before you rely on specific values. See [Benchmarks](/benchmarks).

## Confidence

`confidence` is the difference between the highest and the second-highest probability:

```text theme={null}
confidence = top1_probability - top2_probability
```

For the answer above:

```text theme={null}
returns  = 0.98943    (top 1)
billing  = 0.005342   (top 2)

confidence = 0.98943 - 0.005342 ≈ 0.984
```

`confidence` ranges from 0 to 1:

* **Close to 1**: one option clearly wins.
* **Close to 0**: the two best options are nearly tied.

<Warning>
  `confidence` is **not** the probability that the answer is correct. A `confidence` of 0.6 does not mean the model is right 60% of the time. It means the best option is 0.6 ahead of the runner-up. If you need the model's estimate for the chosen option, read `probabilities[choice]`.
</Warning>

### Why a margin

Krun uses the margin because, in Krun One's evaluation, it separated answerable requests from requests where no option applies as well as the alternatives tested, with the lowest error rate on accepted answers. The [abstention](/concepts/abstention) threshold is applied to this value: when `confidence` is below the threshold, the model abstains.

A margin also catches cases a single probability misses. With three options at `0.48`, `0.47` and `0.05`, the top probability looks moderate, but the margin of `0.01` shows that the model can't tell the first two apart.

## Using the signals

<CodeGroup>
  ```python Python theme={null}
  answer = result.answers["department"]

  # Probability of the chosen option
  if answer.choice is not None:
      p = answer.probabilities[answer.choice]

  # Runner-up, for example to show a second suggestion
  ranked = sorted(answer.probabilities.items(), key=lambda kv: kv[1], reverse=True)
  best, runner_up = ranked[0], ranked[1]
  ```
</CodeGroup>

Recommendations:

* Branch on `choice` and `abstain` first. They already apply the model's threshold.
* If you need a stricter policy, for example before an irreversible action, add your own threshold on `confidence`, tuned on your traffic.
* Log `confidence` and send [feedback](/guides/feedback) so you can see how it relates to errors in your data.