Score accounts with Jev (Jev decides, code counts)

The insight: a lead score is a handful of judgments plus arithmetic, and the arithmetic must never be asked of the model. "Is this a distributor?", "do they run their own fleet?", "how strong is this growth signal?" are judgments. "Are they between 200 and 2,000 staff?", "is the HQ in North America?", "how many points is that, and is it an A?" are arithmetic. An AI column that is asked for "a fit score from 0 to 100" mixes the two and returns a number nobody can decompose, audit or reproduce.

The evidence is Jev's own documentation. TypeSafe publishes a list of what Jev 1.13 is bad at, and the top of it is numbers, counting and comparing dates; it reads instructions literally; it degrades when the record is padded with fields the question does not need. Its recommended pattern for scoring is to break the judgment into atomic questions and combine them with weights in code. And what Jev returns is not a label but a probability for every option, so a score can use how sure it was.

What follows is the whole design: