What Jev is, in plain terms
A judgment model, not a chat model. You give it some text and a set of typed questions; it gives back probabilities. It cannot write you a sentence, and that is the point.
01A model that decides instead of writing
Every model you have used until now produces text. You ask a question, it writes sentences, and your code picks the answer back out of those sentences. Jev removes that round trip. You give it some state and declare the questions you want answered, each with a type, and it returns values of exactly those types.
TypeSafe calls this class of model a System One model, and describes Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out
. The comparison to human cognition is deliberate: fast judgment rather than deliberate reasoning.
02The three questions you can ask
There are three question types, and everything Jev does is built from them. A single request can mix all three, and they are answered together against the same state.
| Type | Ask it | Get back |
|---|---|---|
| Boolean | Is this statement true? | One probability between 0 and 1 |
| Choice | Which of these options fits? | The chosen option, plus a probability for every option |
| Score | Where does this sit on a scale? | An interpolated score, plus the probability of each rung |
03What an answer looks like
Ask several questions about one piece of text and you get several independent judgments in a single round trip. Nothing here was written by the model; every value is a number it assigned.
04What it will not do
The limits are the interesting part, because they are structural rather than temporary. Jev does not generate tokens one after another, so there is no sentence for it to get wrong.
- It cannot write text, rewrite your draft, summarise, translate or hold a conversation.
- It takes text only. No images, audio or video.
- Every question in one request shares the same state.
- TypeSafe states it cannot hallucinate and cannot return a type error, because the output shape is fixed by the question rather than produced by the model.
05Who it is actually for
If a person reads the output, you probably want a chat model. If code reads the output and branches on it, Jev is the better shape: routing support tickets, filtering a feed, grading submissions against a rubric, flagging risk, or picking one of several paths in an agent. It is also cheap enough to run on every item in a queue rather than a sample.
The next page is the catalogue of what people have actually built on it so far, and the one after that has working code for all three ways to reach it.