Task-correct metrics
Choice and Noul use different metrics. The report does not force every task into one generic score.
Jev evaluation
Evaluate Jev on the labels and edge cases you actually own.
A public example cannot tell you how one Jev version handles your classes, questions, and failure costs. Build a dated decision record from your own labeled rows.
An evaluation is more than one accuracy number.
Row-level evidence
Review errors, fallbacks, and class behavior instead of accepting an aggregate without its failure cases.
Versioned context
Record the model route, question hash, dataset hash, threshold method, and measured limitations.
Browse the full method