Jev is an evaluation model. Give it a bounded state and one or more typed questions; it returns choices, scores, and boolean probabilities rather than free-form prose.
One state, independent questions
Every question is evaluated against the same supplied state. A choice picks from named candidates, a score locates the state on an ordered rubric, and a boolean estimates the probability that a proposition is true. Questions do not secretly form a chain where a later answer reads an earlier one.
Three layers that should stay separate
- Evaluation: Jev returns a typed signal and, where supported, a distribution.
- Policy: your code decides whether the signal is strong enough to act on.
- Action: your program performs a bounded operation only after validation.
Why the separation matters
A model may select “fast text” while your policy sends the task to human review because a requested side effect crosses a safety threshold. The model choice remains intact; the application path records the code-owned decision.