Reliability curve
Compare predicted confidence with observed correctness across buckets and keep sample counts visible.
Jev calibration
Confidence only helps when it matches observed outcomes.
Inspect reliability, expected calibration error, and confidence buckets on one recorded dataset before turning a Jev confidence value into automation.
Read confidence as evidence—not certainty.
Task semantics
For Noul, values near zero can represent a confident negative—not weak confidence. The report preserves that distinction.
Small-sample guardrail
Insufficient rows or class coverage trigger EXPLORATORY / INSUFFICIENT_SAMPLE instead of a production recommendation.
Browse the full method