Start here

Jev in 60 seconds.

Jev is TypeSafe’s System One model: it takes application state, answers explicit narrow questions with a typed result and a confidence, and hands the answer back to code that stays in charge of what happens next.

Official Documentation

Three question types

Choice
Picks one option from a list you define, with a probability for each.
Score
Places something on an ordered scale you define.
Noul
Gives the probability that a yes-or-no statement is true.

Example questions and values, drawn to show the shape of each answer. They are not measurements.

The shape of an integration

01Unstructured stateWhatever your application already has: a message, a diff, an event, a screen, a row.
02Narrow independent questionsDecomposed so each one can be answered on its own, and many can be asked at once.
03Choice · Score · NoulThe typed primitives. A bounded option, a calibrated number, or a probability.
04Probabilities and confidenceOrdinary values your code can threshold, log, compare, and regression-test.
05Deterministic application codePolicy, side effects, retries, and anything that has to be reproducible.

The interesting consequence is the last row. Because the model returns a value rather than prose, the decision boundary between the model and your program is explicit — which is what makes the rest of this atlas possible to reason about.

Jev is

A fast, narrow judgment layer that application code calls like any other function.
Typed output, so there is no free-form response to parse or repair.
Cheap enough per call that judging every item in a stream becomes affordable.
A source of probabilities that deterministic code composes into policy.

Jev is not

Not a chat interface. There is no conversation, persona, or message history to manage.
Not a code generator. It answers questions; it does not write your implementation.
Not autonomous control flow. It never decides what runs next — your code does.
Not a replacement for deterministic rules. Anything that must be exact stays in code.
Not a correctness guarantee. A typed answer constrains the shape of a decision, not its truth.

What is actually established

This atlas separates what has been demonstrated from what TypeSafe says about its own product. Both matter; they are not the same kind of statement.

Demonstrated · 3

Jev exposes constrained decision primitives for Boolean probability, choice distributions, and ordered scores.
Typed output removes free-form parsing but does not make wrong decisions impossible.
Developers have published small Jev integrations and demonstrations.

Vendor claim · 2

TypeSafe reports Jev as materially faster than LLM workflows on its own evaluations.
TypeSafe reports a low input-token price and no metered output-token charge for Jev.

Open the full claims ledger

Where to go next

Opportunity mapThe design space, grouped by the property that makes Jev useful.Architecture patternsThe structures developers keep converging on, with caveats.Case studiesWork that was actually located, kept separate from proposals.For agentsMCP endpoint, JSON exports, and llms.txt for coding agents.

Official sources

TypeSafe’s documentation is the authority on the current API contract. This page explains where Jev fits; it does not restate the API.

Three question types

Choice

Score

Noul

The shape of an integration

Jev is

Jev is not

What is actually established

Where to go next

Official sources

Three question types

Choice

Score

Noul

The shape of an integration

Jev is

Jev is not

What is actually established

Where to go next

Official sources