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

# Decisions

> The three decision types — route, score, classify — and how to define them.

A decision is a closed question your workflow asks Branch Pilot. You define it in the editor: what the workflow sends (the **context**), the **question** asked to the engine, and the **options** it can answer with.

## Context

Describe what your workflow will send, and optionally declare fields:

| Setting             | Effect                                                                                                                                                                          |
| ------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Field name and type | The payload is validated; a missing required field or a wrong type returns `400 invalid_payload` with the problems listed.                                                      |
| **PII**             | The field is replaced by a token before any model call, and its value is also replaced wherever it appears in free text. See [pseudonymization](/en/concepts/pseudonymization). |
| Example payload     | Pre-fills the playground and the test-case dialog.                                                                                                                              |

If you declare no field, any payload is accepted as is: an object, a string, or an array.

## Route

Picks **one** option among 2 to 255. Each option has a key, returned to your workflow, and a description the engine evaluates.

```text theme={null}
sales      Wants to buy, asks for pricing, a quote, a demo or a trial.
support    Existing customer with a problem, a bug, a login or billing issue.
spam       Unsolicited offers, SEO or link building, irrelevant content.
```

Response: `choice`, `probabilities` (one per option, summing to 1) and `confidence`.

<Tip>Add an `other` or `none` option when the list may not cover every input. It gives the engine somewhere to put ambiguous cases instead of forcing a wrong branch.</Tip>

## Score

Returns a **0–100 score** over 2 to 10 ordered levels you describe, from lowest to highest.

```text theme={null}
0  Cold       Curious, student, no budget, vague or off-topic.
1  Lukewarm   Interested but no timeline, small team or unclear need.
2  Warm       Clear need, relevant role, mentions timeline or team.
3  Hot        Decision maker with budget, urgency and a concrete use case.
```

The engine returns a probability per level. `raw_score` is the weighted position on the level scale (for example 2.4), `score` is that position remapped to 0–100, and `legend` maps level indexes to their labels.

## Classify

Returns **every label** whose probability reaches its threshold — several at once, or none. Each label has a key, a description and an optional threshold (default 0.5, adjustable per decision).

```text theme={null}
billing          Invoices, payments, charges, refunds or plans.
technical        A bug, an error, something not working.
feature_request  Asks for a new feature or improvement.
```

Response: `labels` (the retained keys) and `probabilities` (one per label, independent of each other).

## Behavior settings

These apply to every version of the decision and take effect immediately:

* **Confidence threshold** (default 0.70) and **uncertainty policy** — see [confidence](/en/concepts/confidence).
* **Engine**: Jev with automatic LLM fallback (default), Jev only, or LLM only.
* **PII detection** in free text: off, standard or strict.
* **Engine timeout** (default 3 000 ms).