Define labels first
Write the expected decision and class criteria independently of Jev's answer.
How to benchmark Jev
A reproducible Jev benchmark starts before the model call.
Define the task, map only necessary fields, inspect dataset coverage, record the route and version, and keep every failure state visible.
Six checks before you trust the result.
Measure the right task
Use Choice metrics for class decisions and two-sided probability metrics for Noul.
Keep a failure record
Preserve retries, missing usage, row errors, PARTIAL states, and provider route changes.
Browse the full method