Run a question set once against a single state and see the typed answers with their probability distributions. Use it to prototype question wording before committing it to code or to jev_eval. Questions over the same state are answered independently in one request, so ask several at once.

Judge every item in a list and get back one compact row per item, plus per-question aggregates. Built for triage and ranking at a scale where reading each item into context is the expensive part: filter, sort, or write the full results to disk and only pull back what matters.

Score one or more wordings of a question against labeled items and report accuracy, calibration (Brier, ECE, AUC), a full threshold sweep for noul questions, an abstention curve for choice and score questions, and the worst misses. Run this before hard-coding any threshold, and rerun it after changing wording or model version. Variants are asked in the same request, so comparing several is nearly free.