Janus

# Janus

[](https://github.com/FirasSX914/Janus/actions/workflows/tests.yml)
[](https://pypi.org/project/janus-decide/)
[](https://github.com/FirasSX914/Janus/blob/main/LICENSE)
[](https://pypi.org/project/janus-decide/)

Janus sends each decision to a small model or to a larger one, according to how
confident the small model is. It measures where that line sits on your data
before it routes anything. **Janus ships no default threshold: it measures one.**

It measures from either of two inputs:

- **a labelled dataset** — `--dataset`, with a gold label, so the report is
  about how often each model is right;
- **a log of decisions already taken** — `--log`, with a confidence and no gold
  label, so the report is about how often the logged model **agrees with a
  reference model**. There is no ground truth in a log, and Janus refuses to
  print a word that would suggest one.

[](https://github.com/FirasSX914/Janus/blob/main/RESEARCH.md)

<sub>Real output, replayed from the raw JSONL committed in this repository. No model is called.</sub>

```bash
pip install janus-decide
```

## Quickstart

Five minutes, on your own data. **Janus ships no default threshold: it measures one.**

**1. Install**

```bash
pip install "janus-decide[typesafe,deepseek]"
```

The bare package depends only on numpy; each backend is an extra. DeepSeek speaks the
OpenAI protocol, so `[deepseek]` and `[openai]` pull the same client.

**2. Prepare a labelled JSONL** — one object per line, three fields:

```json
{"id": 0, "text": "I lost my card", "gold_label": "lost_or_stolen_card"}
{"id": 1, "text": "when does my card arrive", "gold_label": "card_arrival"}
```

and a label file naming every class you allow:

```json
{
  "instructions": "Which banking intent does this customer query express?",
  "labe