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

# Results

> How thresholds turn model answers into accept or review.

The model returns answers. The pack's `review` rules decide the status. Any fired rule stores `review`. No fired rule stores `accept`.

`review` on the question is yours. It is not shown to the model.

## Choice

Review when the selected option has no probability, or that probability is **below** `minSelectedProbability`.

A probability equal to the floor is `accept`. A missing `probabilities` object is `review`, with reason `{id}: selected probability is missing`.

## Score

Review when `score` is inside `[minInclusive, maxInclusive]`, including both ends.

Reason: `{id}: score is in the review range`.

## Boolean

Review when `lowExclusive < probability < highExclusive`. The ends are outside the band, so they are `accept`.

Reason: `{id}: probability is in the review band`.

A boolean answer has no separate confidence field. `probability` is the evidence.

## Invalid answers

An answer with the wrong type for that question is `review`, with reason `{id}: invalid answer`.

## Model and gateway failures

A provider failure or invalid model output is stored as `review`. `answers` is empty. `modelId` is `none`. `reviewReasons` contains one of:

| Reason | When |
| - | - |
| `gateway_auth` | The gateway rejected the credentials (`401` or `403`) |
| `gateway_unavailable` | The gateway could not be reached |
| `invalid_model_output` | The model output did not match the questions |
| `invalid_pack` | The pack document failed validation at judge time |
| `invalid_ticket` | The ticket failed validation inside the kernel |

The HTTP call still returns `200` with that stored result, except when the request itself is rejected. See [Errors](/docs/guides/errors). A gateway failure does not consume an idempotency key.

## What the caller receives

```json theme={null}
{
  "status": "review",
  "answers": {},
  "reviewReasons": ["urgency: score is in the review range"],
  "modelId": "typesafe-ai/jev",
  "latencyMs": 840,
  "usage": { "inputTokens": 900, "outputTokens": 120 },
  "derived": { "department": "billing", "priority": "medium", "tags": ["refund"] }
}
```

`usage` and `derived` are omitted when they are empty. `status` may also be `reject` in the schema. Current threshold rules do not produce it.

Choice and score confidence, when the provider sends it, lives in provider metadata on the stored evaluation (`providerMetadata.typesafe.confidence`). It is not an extra cutoff. Boolean answers do not have that field.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.