Jev

Jev is an AI model that makes decisions instead of writing text. You hand it a block of information plus questions whose possible answers you define up front, and it hands back typed answers with a probability attached to each. It is the first release in a category TypeSafe AI calls System One models, in early access since September 2026.

What It Is

Every call has two parts: a state, the text or JSON being judged, and typed questions about it. Choice picks one option from a labeled list. Score places the state on an ordered scale you describe in plain language. Noul answers a yes-or-no statement as a probability between zero and one. A language model produces one token at a time, each depending on the last, even for a small JSON object. Jev has no autoregressive step at all: a parallel sampler answers every question in one forward pass against the same state, which TypeSafe says yields responses in the low hundreds of milliseconds however many questions you attach. Because the answer space is declared in advance, malformed output is impossible by construction, though a type-correct answer can still be wrong. Every answer also carries a confidence value, and that is the number your code uses: act on the confident ones, escalate the rest.

ELI5 - What's Jev and What's TypeSafe

TypeSafe is the company. Jev is the model it built. System One is the category it coined for models that work this way. Picture an ordinary chatbot as a colleague who answers everything with a paragraph you then have to interpret. Jev is the colleague who only ever fills in the form you handed over: tick one box, give a rating, answer yes or no, and mark how sure they were. Developers landed on a shorter version, a very smart switch statement. TypeSafe makes the point at the door, gating early access behind a one-question quiz: can you chat with Jev? Answering no is what lets you in, and the reveal runs that same question back through the model as a Noul, landing at 97 percent false.

Where the Name Comes From

Founder Diogo Almeida wrote in his launch thread that Jev is named after William Stanley Jevons, the nineteenth-century economist behind the Jevons paradox: when coal-burning steam engines became more efficient, Britain burned more coal, not less. The bet in the name: once a decision gets cheap enough, software starts making vastly more of them. The category name borrows from Daniel Kahneman's Thinking, Fast and Slow: System 1 is fast and intuitive, System 2 slow and deliberate.

How You Build With It

TypeSafe ships an official agent skill for Claude Code, Codex, and other coding agents, loading the question types and patterns into the agent's context. Its suggested opening prompt is the clearest statement of where Jev belongs: point the agent at your project and ask it to find places where judgment could stand in for complex parsing or other fragile code. The docs recommend keeping your questions and thresholds in one file, since those constants are what a human needs to review, and warn that agents write poor questions.

What People Have Built With It

Developers quickly wired Jev into places an LLM is too slow to sit. One build has it playing Doom at roughly ten decisions per second. Browser Use published an open-source browser agent that finished a flight search in about seven seconds, asking Jev to pick the next action from the live page. A researcher classified 1,018 AI papers into topic buckets for a research explorer. Others built instant context compaction, scoring every tool call in an agent's history and dropping irrelevant ones, plus a trading bot and a drone stack that call it inside real-time loops. The shared pattern already has a community name, Jev Engineering: an LLM writes, Jev decides, ordinary code acts.

Limits Worth Knowing

Jev cannot write text, code, or summaries, so it complements a frontier model rather than replacing one. TypeSafe documents that it reads state literally, is unreliable at arithmetic, date math, and counting, and degrades when the state carries irrelevant material. A single Choice question tops out at 255 options, and input is text only. TypeSafe also publishes a known-failure-modes page and notes its benchmarks are self-run.