Jev primer: moving yes/no decisions out of the LLM
Introduces Typesafe's Jev, which takes state and typed questions and returns only probabilities, not sentences, at $0.042 per million input tokens, and covers its three decision shapes and criteria for migrating calls to it.
I still don't understand why we pay a language model to answer yes or no. six calls in my last agent session, and none of them needed a sentence. TypeSafe AI launched Jev on September 15, 2026. it can't write a word. state and typed questions go in, a probability for every valid answer comes back. $0.042 per million input tokens, output free. ten thousand decisions cost 42 cents. three questions decide what moves over. can you list every valid answer? would a person settle it in a glance? does it fire often enough for milliseconds to matter? three yeses, the call leaves the LLM. one no, it stays. three shapes cover everything that moves. Choice picks from up to 255 options. Score places the input on a scale and can land between two levels. Noul returns the probability of yes. code computes what it can. Jev decides what only judgment can. code executes the decision. nobody asks a model whether the tests passed, they read the exit code. a high number doesn't mean a right answer. the model can be sure and wrong, and the score looks identical. set a threshold per decision type and send everything under it to a stronger model. most steps in an agent were never asked to be cheap. do this today: open your last agent run, mark every call that has a fixed list of answers, run each one through the three questions. the ones that pass are your first candidates. one rule at every scale. if code already solves it, keep the code.
This post quotes another post: x.com/imryven/status/2104255840819491030