What is Jev?

How a call is shaped

Every call has two parts: the state, which is the context to judge, and the questions you want answered about it. Questions come in three kinds and can be mixed in one call. Each is evaluated in parallel against the same state, so asking sixty questions costs about what asking one does.

The probabilities say which option wins. The confidence summarises how concentrated that distribution is, and that is the number your code acts on: handle the confident cases automatically, send the rest to a person or to a larger model.

Speed, cost and accuracy

Input costs $0.042 per million tokens and output is not charged. Responses land in 70 to 500 milliseconds. TypeSafe's own workflow evals put it level with a mid-sized reasoning model on accuracy while being about two orders of magnitude cheaper per case.

Source: TypeSafe workflow evals. Read the accuracy column carefully: the reference answers are the average of two large reasoning models, not a human-labelled set, so 67.8% means “agrees with those two 67.8% of the time”.

Where it fits

The projects in this directory cluster around four patterns: picking the next action in an agent loop, sorting a queue into buckets, scoring every item in a corpus, and gating a risky step behind a confidence threshold. All four are decisions a system makes thousands of times, where the cost of a slow answer is the whole problem.

Where it does not fit

• It writes nothing. Drafting, summarising and code generation stay with an LLM.
• The answer space is yours to define. A Choice takes at most 255 options.
• Text only, for now: strings, JSON objects and lists of strings. No images, audio or video.
• “Cannot hallucinate” means the shape is guaranteed, not the content. Measure calibration on your own data before you set a threshold.

Common questions

What is a System One model?

A model class trained for fast, bounded decisions that software consumes directly, named after the fast, intuitive System 1 in Daniel Kahneman’s Thinking, Fast and Slow. It does not write prose, generate code or explain its reasoning.

Can Jev hallucinate?

It cannot return a value outside the type you asked for, so a broken JSON payload or an invented field is impossible. A well-typed answer can still be the wrong answer, which is what the confidence score is for.

What does Jev cost?

TypeSafe prices input at $0.042 per million tokens and does not charge for output. End-to-end response time is 70 to 500 milliseconds.

Does Jev replace an LLM?

No. It replaces the small, repeated decisions inside a system — classify, route, score, filter, gate a tool call — while a larger model still handles anything that has to be written.