TypeSafe’s build guidance separates the two halves. Ask many independent parallel questions about the same state, then “compose answers with deterministic rules or code-controlled weights”, keeping deterministic work in code. That split matters here, because arithmetic and counting are on jev-1.13’s published list of weak spots. You do not want the model doing the summing.
A lead-scoring example. Ask four Score questions against the same state, each with its own ordered levels: budget fit, urgency, decision-making authority, technical fit. Each returns a probability-weighted mean of the level numbers, the way the docs’ bug-severity example returns 1.30 from probabilities of 0.0, 0.70 and 0.30 over levels 0, 1 and 2. Your code multiplies each by a weight you chose and adds them up.
The weights are yours to tune and to audit. When a deal is misrouted you can point at which number moved, and you can change the weighting without touching the model at all. Confidence comes back per question, so a dimension that scored below your floor can be dropped or sent for review before the combination step.
See the composite lead scoring recipe and the score primitive.