Open Jev (intikhab49)

# Open Jev — a typed decision engine you can train for free

**A 150M encoder that answers arbitrary typed questions about a state in one forward pass, with calibrated confidence. 0.03 behind TypeSafe Jev on its own benchmark, 2.5× better calibrated, 4× faster, $0.**

[](LICENSE)
[](https://www.python.org/)
[](jevlite_colab.ipynb)
[](https://huggingface.co/datasets/LocalLLaMA/typed-decisions)

[Results](#results) · [How it works](#how-it-works) · [Quick start](#quick-start) · [What we learned](#what-the-runs-actually-proved) · [Limitations](#limitations)

</div>

---

## What this is

[TypeSafe AI's **Jev**](https://www.mindstudio.ai/blog/jev-system-one-model-launch) (launched 2026-09-15) is a "System One" model: it never writes prose, it only **decides, classifies, routes and scores**. You hand it a state plus a set of typed questions and it answers all of them in one non-autoregressive pass, with a confidence on each. $0.042/1M input, output tokens free, ~239 ms per call.

This repo reproduces that interface with an open 150M encoder you can train on a free Colab T4 in under 30 minutes, then run locally — or in a browser — for nothing.

> **TypeSafe does not publish Jev's parameter count.** This model is 150M. Their [published specs](https://docs.typesafe.ai/models) give pricing, rate limits and a 64k context window, but no model size — and their launch FAQ lists *"Is Jev just a smaller LLM?"* without answering it. Any size comparison you see, including here, is inference from price and latency rather than a disclosed figure.

Three primitives, matching theirs:

| type | meaning | criteria format |
|---|---|---|
| `noul` | boolean | `{"true": desc, "false": desc}`, or omitted entirely |
| `choice` | enum | `{label: desc, ...}` |
| `score` | ordered scale | `[level_0,