handbooksWhat is Jev?
Jev is a model from TypeSafe AI that returns decisions, not text. You give it evidence and a few typed questions, and it gives back choices, scores, and probabilities your code can use directly.
- Jev is a decision model. It does not write prose. It picks from answers you define and tells you how sure it is.
- You send state (the evidence) and questions (what to decide). You get back typed answers with probabilities.
- Three question types: Choice (pick one option), Score (place it on a scale), Noul (yes or no, as a probability).
- It is fast and cheap: 70 to 500 ms per call, with output tokens free.
- Use it for routing, classification, scoring, and guardrails. Keep the LLM for anything that has to write words.
The idea in one picture
Most AI features today ask a chat model a question and then parse whatever text comes back. That works, but it is slow, it costs output tokens, and sometimes the text does not match what your code expected.
Jev flips it around. You list the possible answers up front. The model can only return one of them.
Why it is called a "System One" model
The name comes from Daniel Kahneman's two modes of thinking. System One is the fast, intuitive snap judgment. System Two is the slow, careful reasoning.
Chat LLMs are built for System Two work: explaining, planning, writing. Jev is built for System One work: "which team owns this?", "is this urgent?", "how bad is it?". Questions a person would answer in a second, but that your software asks thousands of times a day.
TypeSafe trains it with a method they call reinforcement learning for calibrated decisions. The goal is honest probabilities. When Jev says 0.9, it should be right about nine times out of ten.
What you get from it
| Property | What it means for you |
|---|---|
| Typed output | The answer is always one of your options. No parsing, no "hallucinated" labels. |
| Probabilities | Every answer comes with a number you can put in an if statement. |
| Parallel questions | Ten questions in one request take about as long as one. |
| Low cost | Input tokens only. Output is free. |
Jev guarantees the shape of the answer, not that the answer is right. It can still pick the wrong team. That is why you test it on real data and use confidence to decide when a human should check.
When to use it, and when not to
Good fit:
- Routing a request to the right agent, team, or model
- Classifying tickets, emails, documents, or search results
- Scoring things on a rubric (severity, relevance, sentiment)
- Guardrails: "is this tool call safe?", "does this message contain PII?"
Not a fit:
- Writing replies, summaries, or code. Use an LLM.
- Math, counting, or comparing dates. Do it in code.
- Images or audio. Jev reads text only, so convert them to text first.
What this handbook covers
| Page | What it answers |
|---|---|
| Your first call | HTTP, Python, and TypeScript, in five minutes |
| Choice, Score, Noul | Which question type to use, and what each returns |
| Writing state & questions | How to phrase things so Jev gets them right |
| Confidence & thresholds | Turning probabilities into safe decisions |
| LangGraph & LangChain | Jev as a router node, and as agent middleware |
| Vercel AI SDK | experimental_evaluate and the AI Gateway |
| Common patterns | Intent routing, fan-out, composite scoring, guardrails |
| Limits, pricing & weak spots | The numbers, and where Jev gets things wrong |
Where this page's facts come from▾
TypeSafe AI's documentation at docs.typesafe.ai, their launch post, Vercel's What is Jev?, and LangChain's Building a harness with Jev. Numbers are for jev-1.13.0, the current model at the time of review.