What System One is

The point

LLMs write text for humans to read. When your code needs a decision, digging an answer out of free-form text is the wrong tool.

Jev is TypeSafe's flagship System One model. You send it state plus typed questions. You get back structured answers and probabilities. Your software can branch on them, sort them, and route with them. No text generation. No parsing.

Here is a simple way to picture it. An LLM writes you a letter. Jev fills out a form.

How Jev compares to an LLM

• An LLM gives you generated text. Jev gives you typed decisions plus probabilities.
• With an LLM, you write a prompt, then parse JSON out of the reply. With Jev, the answer is structured from the start.
• LLMs fit chat, generation, and open-ended reasoning. Jev fits fast, narrow judgment calls inside a workflow.

What Jev does not do

Jev does not write replies, generate code, or explain its reasoning. You define the space of possible answers yourself, using the primitives: Choice (pick one option from a set), Noul (yes or no, with a probability), and Score (rate something on a fixed scale).

What goes in

Input is text only. That can be a plain string, a JSON object, or an array of text. Images, audio, and video are not supported yet. English is the primary language.

Name note: "System One" echoes Kahneman's fast System 1. The emphasis is on focused judgments. Details in System One.

Training framing from TypeSafe's docs: RLCD means reinforcement learning for calibrated decisions. Generative models use RLHF or RLVR instead. More in the AI primer.

Next

Move on to When to use Jev, an LLM, or plain code. Then verify your key.