Jev beats llama3:8b 18/20 to 15/20 on store-diagnosis test

Jev is a different kind of model from a chatbot: TypeSafe AI's "System One" model, released on September 15, 2026, does not generate text at all. You hand it an app's state (text or JSON) and a typed question, and it returns an answer drawn only from a predefined set of choices, plus probabilities and a confidence value. The question types are Choice, Score, and Noul. Because the answer must come from the defined options, it cannot invent a nonexistent answer — though it can still pick a wrong one.

The author's step-by-step experiment shows why input design moved the result more than the model did. When a prior-month rainy-day count was missing, Jev leaned toward weather at 0.65 and menu at 0.29; adding that one line pushed weather to 0.99 — but it also crushed the menu cause to 0.01. Because Choice asks for a single main cause, a second cause gets starved of probability. Asking about each cause separately with Noul brought the menu cause back up to 0.41. In a later round, the menu score fell again until the author added a menu_changed_this_month flag; the phrase "a major factor" versus "one of the factors" barely moved the number. The underlying problem was that the data recorded a drop in spend per customer but never stated that the menu had changed — the model could not infer a cause its input did not support. One acknowledged limitation: the change-rate inputs and the rewritten question were altered in the same round, so their individual effects are not separated.