Jev: Choice, Score and Noul
Jev evaluates typed questions against shared state. Use its structured answers to support routing, scoring and verification; validate accuracy on your own workload before automating actions.
- •Choose a typed primitive for a decision.
- •Separate independent questions from dependent workflow steps.
- •Evaluate decision quality and choose an escalation threshold.
Sources & further reading
TypeSafe AI: introduction and quickstart ↗In-Depth Architecture & Mechanics
Choose the question type
Choice selects from options; Score assesses a rubric; Noul represents the probability that a statement is true. Choice and Score also expose confidence. Keep each question narrow enough to evaluate one factor.
Compose the workflow
Independent questions can be evaluated against the same state in one request. If a later question requires an earlier result, combine results in application code or make a subsequent request. A typed response makes integration simpler but does not guarantee a correct judgment.
Professional lab: support triage
Create a labeled set of realistic support tickets with an intent, urgency and expected route. Include ambiguous tickets and out-of-domain examples. Define explicit choices for intent, a documented scoring rubric for severity, and a separate true-or-false question for whether evidence is sufficient. Use the official quickstart for the current API schema; keep credentials in a server-side environment.
Decide when to abstain
Measure per-class precision and recall, calibration and review volume on held-out examples. Select a threshold using the cost of incorrect routing and human review. Route missing, invalid or uncertain results to review; log model version and question definitions so you can compare changes. Do not interpret confidence as guaranteed correctness.