Browse published operations

Judges state against one yes/no question (Jev's "noul" type) and returns a 0-1 probability -- no generated text. Exactly one question per call (a single key under questions, typed noul); pass criteria.true/criteria.false as concrete situations to improve calibration, not adjectives. Reads state literally: spell out edge cases in instructions, do not rely on the model inferring them. Pin model to jev-1.13.0 unless told otherwise.

Judges state against one score question (Jev's "score" type) and returns a probability-weighted position along an ordered rubric (can land between levels, e.g. 1.6), with the legend and a confidence -- no generated text. Exactly one question per call, 2-10 rubric levels written as concrete situations (not adjectives). Score is for thresholds and ranking, not an exact interpolated number -- do not treat it as ground truth to two decimal places. Pin model to jev-1.13.0 unless told otherwise.

Judges state against one choice question (Jev's "choice" type) and returns the winning option, a probability per option, and a confidence -- no generated text. Exactly one question per call, 2-20 options. Reads state literally and is not a calculator: keep counting/ranking logic in your own code, ask Jev only to judge semantic fit. Pin model to jev-1.13.0 unless told otherwise.