How the Jev architecture works

The Jev architecture is best understood from the outside in: Jev returns typed, calibrated decisions instead of text. This page describes the observable Jev architecture, since TypeSafe AI has not publicly disclosed the model's internal weights or parameter count.

Jev is TypeSafe AI's first System One model, and the Jev architecture reflects that framing. Where a frontier LLM writes fluent prose, the Jev architecture is designed to answer questions with structured, calibrated outputs you can act on directly. Every response is locked to a type you defined, so the observable Jev architecture behaves less like a chatbot and more like a fast decision engine.

The request and response shape

A single call to the Jev architecture is made of three things: a model, a shared state, and a map of questions. The state is the context every question sees, and each entry in the question map is answered independently. Answers come back keyed to the same map, so the shape of the response mirrors the shape of your request. This request-and-response symmetry is the most visible part of the Jev architecture.

Three typed primitives

The Jev architecture exposes exactly three decision primitives, and every question uses one of them:

• choice: pick one of up to 255 options, returning the chosen key plus per-option probabilities and a confidence.
• score: rate on an ordered level scale, returning a fractional score, a distribution and a confidence.
• noul: a calibrated yes/no returned as a probability from 0 to 1, with optional true/false criteria descriptions.

Because outputs are typed, the Jev architecture cannot return an option you did not list or a malformed shape. That type-locking is what TypeSafe means by zero hallucination in the Jev architecture.

Parallel evaluation and calibration

Questions in a call run in parallel in one round trip, each evaluated in isolation against the same shared state. This parallel, isolated design lets the Jev architecture stay fast, roughly 40 to 200 times faster than a frontier LLM. Every answer also carries a probability distribution and a confidence, produced by the model's RLCD training objective. Calibration, not fluent writing, is the target of the Jev architecture.

What is public and what is not

Be clear-eyed: the Jev architecture is a hosted API, and TypeSafe has not published layer counts, parameter counts or low-level internals. What you can rely on is the contract described above. Treat confidence as how concentrated an answer is, not a correctness guarantee, and calibrate thresholds against your own labelled data when you build on the Jev architecture.

See also: What is Jev · How to use the API · Playground

Related: RLCD · System One model · Jev vs an LLM