Jev Is How Judgment Becomes Infrastructure
A decision dressed up like a primitive
Jev is built by TypeSafe AI as "System One" — a model from Diogo Almeida's team (formerly of OpenAI) that does not generate text at all. Because it cannot generate text, it cannot hallucinate text. You declare a return type, and it returns one of three typed values with calibrated confidence:
| Jev primitive | What it returns | The code it replaces | Example |
|---|---|---|---|
| Choice | one label from a declared enum (up to 255) plus the full probability distribution | switch / match on an enum | route an inbound request to one of N handlers |
| Noul | P(yes) between 0 and 1 | if | is this function destructive? |
| Score | an ordered value on a rubric | comparator, sort key, threshold | rate this feedback 1–5 |
switch/match statement, correlating to the enum you switch on — the enum choice being the important part. A noul correlates to an if. A score correlates to sorting or thresholding. It is a familiar paradigm, so engineers get it at a glance. The stated intent is that more types will come and correlate the same way. The interface is a type, and the output is a decision, not prose to parse.It returns a decision — typed, calibrated, in a single pass — in roughly 100–500 milliseconds, at a cost of about four cents per million input tokens (output is free). Fast and cheap enough to sit inline in code.
A couple of example use cases
A couple, to give a feel for calling it: in an agent's Ask mode, a single noul call decides whether a function — a raw string of code — is destructive, and gates execution in milliseconds.
Built for engineers, not for chat
Jev is not a chatbot, and it does not want to be. It is built for people who write code, to be called from code, and to return values code can use. There is no prose to read, no conversation to manage, no persona. The interface is a type. The output is a decision. That is infrastructure: something you import like a library and let disappear into your control flow.
Faster and cheaper would be a step change; a model whose outputs are programming primitives is a category change, and I expect this to be remembered as a defining moment in our agentic progression.
Sources
- TypeSafe AI — Introducing System One Models and Jev (launch post): https://www.typesafe.ai/blog/introducing-system-one-models-and-jev
- Latent Space — Why We Made Jev (Diogo Almeida, TypeSafe co-founder & CEO): https://www.latent.space/p/jev
- TechCrunch — coverage of TypeSafe and Jev (September 2026): https://techcrunch.com/2026/09/18/typesafe-jev/