What is Jev?
A short, plain explanation of System One models, the three question types, and the ways to reach the model.
A chat model answers by writing. That is useful when a person reads the answer, and awkward when a program has to act on it: the text has to be parsed, the parse can fail, and a value that was never one of your options can still come back. A System One model removes that step. You describe the decision as a typed question, and the model returns a value from inside that type.
The shape of a call
You send state, not a prompt
The request carries the thing being judged — a string, an object, a list — under its own field, separate from the questions. It is sent once, however many questions you attach to it.
You attach named, typed questions
Each question has a name you choose, a type, and instructions. The answers come back under the same names, so there is no positional matching to get wrong.
Every question is answered in one pass
Batching is not an optimisation you bolt on later; it is the normal shape of a call. Asking ten questions about one state costs roughly one state, not ten.
The answer arrives with a probability
The value is what to do; the probability is how much to trust it. Thresholds live in your code, which means the risk you are willing to take is reviewable in a diff rather than buried in a prompt.
The three question types
choiceReturns: one option, plus the probability of each
Pick one label from a set you defined. Use it for routing, categorising and any decision whose outcomes you can enumerate.
scoreReturns: a probability-weighted level on your scale
Place the state on an ordered rubric. Use it when the levels have a direction — severity, quality, priority — rather than being unrelated buckets.
noulReturns: a single probability between 0 and 1
Ask whether something holds. There is no separate confidence field, because the value already is the confidence.
Where you can reach it
Besides the official API there are relays you can reach with a key you already have. If you want it on your own hardware, there are open implementations that speak the same protocol.
- Jev on Vercel AI Gateway — Vercel
- LiteLLM TypeSafe pass-through — BerriAI (LiteLLM)
- LocalJev — GitHub Next
- OpenJev — razorback16