A JEV question is not “what do you think?”. It is “answer inside the shape I gave you.” The public write-ups collapse that shape into three types. Picking the type matters more than writing a longer prompt.
choice: pick one option from a closed list. Support topics billing / account / technical, agent actions allow / confirm / block, content labels ok / spam / abuse — all choice. The options are fixed when you ask. The model selects; it does not invent a fourth class.
score: return an ordered level. Urgency low / medium / high, plus risk, severity and priority. The difference from choice is that order means something — high is stronger than medium — so you can attach a different policy to each rung.
noul: return a probability from 0 to 1. Needs a human? Is the tool dangerous? Should this come down? The point of noul is calibration: 0.93 and 0.41 can drive different workflows. The sentence “I’m a bit unsure” cannot.
One request can carry several questions over the same state. Ask topic, urgency and human-need on one support message. Then wire thresholds: p ≥ 0.85 acts, 0.50–0.85 confirms, below 0.50 escalates. Types give you shape. Probability gives you nerve.