Calibration
Calibration — a probability (or confidence) is calibrated if events with reported p ≈ 0.8 happen about 80% of the time.
TypeSafe’s launch post says Jev is trained with RLCD for “epistemically honest probabilities” and that higher confidence means higher accuracy. That is a vendor claim. Official confidence docs: confidence is a convenience statistic of probabilities; you may use another. Noul has no confidence field.
Measure ECE / reliability diagrams on your labels (offline evaluation). Do not treat launch-post plots as your ECE.
See also: confidence, confidence threshold.
Unofficial. Confirm docs.typesafe.ai/confidence.
Sources
Public TypeSafe or adjacent documentation only. No private claims.