What Is Jev

This OpenRouter explainer answers what a decision model is and how to read what it returns, using Jev from TypeSafe as the example (openrouter.ai). A decision model reads text, and in Jev's latest version images, but never writes any. The developer supplies the input plus a typed question, and the model answers only from options defined in advance. Jev handles four question types: Choice picks one option from a set and gives a probability for each; Score places the input on ordered levels the developer describes; Noul returns the probability that a yes-or-no statement is true; and Box, which works on images, returns the edges of a box around a described target. Because nothing free-form comes back, there is no text to parse. The post also lists what Jev leaves to the developer: it writes no explanations or reasoning, makes no tool calls or multi-step plans, and is not meant for exact arithmetic or date math, which should be computed first.

The useful part is how to read the numbers. TypeSafe calibrated Jev so that answers given at 0.8 are right about 80 percent of the time, but the post stresses that this holds only on average across many answers; any single one can still be wrong. A Noul near 0.5 means the model is genuinely unsure, and the author suggests treating it as a third outcome to follow up on or hand to a person. The confidence figure on Choice and Score answers measures how concentrated the probabilities are, not whether the chosen option is correct. Box always returns a box, even when the target is absent, so the post suggests asking a Noul about the target in the same request. Thresholds should come from a few hundred of a team's own labeled examples. Jev is proprietary, with no published weights or paper, and as of October 16 OpenRouter listed 21 decision models from 13 makers behind the same Decisions API. A companion post shows Jev and an LLM working together on support tickets.