What Is Jev? TypeSafe System One AI Explained
Understand Jev’s typed decision interface, what it returns, and when to use it beside an LLM.
On this page
What Jev doesWhat “System One” meansWhat goes in and what comes outWhere Jev fits beside an LLMCurrent version, price, and limitsHow to startJev is TypeSafe AI’s first System One model. Instead of generating open-ended text, it evaluates state against typed questions and returns bounded Choice, Score, or Noul decisions with probabilities.
What Jev does
A support application might need to identify the right department, rate urgency, and detect whether a refund was requested. Jev can evaluate all three against one ticket. Your code receives structured results and decides what happens next.
The important boundary is the answer space. You supply the categories, rubric levels, or yes/no proposition. You do not ask Jev to invent an explanation or compose a customer reply.
What “System One” means
TypeSafe uses this term for fast, structured judgments made inside software. It is a product and interface framing, not a public specification of the model’s internal architecture. This manual explains observable API behavior rather than guessing the training architecture.
What goes in and what comes out
The input is state: a string, JSON object, or array. Each question has instructions and, where required, criteria. The output has a versioned model identifier, one answer per question key, and token usage.
| Need | Primitive | Example |
|---|---|---|
| One category | Choice | billing, technical, sales, other |
| An ordered level | Score | routine through immediate harm |
| Probability of yes | Noul | customer explicitly requests a refund |
Choice and Score include distributions and confidence. Noul returns a number from zero to one without a separate confidence field. A valid type does not guarantee the correct judgment.
Where Jev fits beside an LLM
Need generated text? Use an LLM. Need one bounded semantic decision? Jev may fit. Need exact arithmetic, identifiers, or a deadline comparison? Use code. A common application asks Jev which handler should receive the task, then lets a generative model write the answer only when needed.
Good early projects include support routing, relevance filtering, source-passage verification, and intent classification. Avoid starting with unsupervised destructive actions: first measure decisions and review mistakes.
Current version, price, and limits
The current direct model is jev-1.13.0. jev-latest and jev-preview currently resolve to this version. Direct input costs $0.042 per million tokens and output is free. These facts were checked on 2026-09-21; provider pricing and identifiers are separate.
Jev accepts text, not native image, audio, or video input. Its reported request budget is 64000 tokens, with 32000 for state plus the longest question. English is the primary training language; evaluate other languages using your own examples.
How to start
Read the interface walkthrough and make a first request. Keep the initial application read-only. Log what it would have done, label mistakes, then introduce a confidence gate backed by those observations.