ecosystem Local OpenJev probabilities look decisive, but TypeSafe Jev (or humans) disagree on the same rows. Updated Sep 19, 2026
OpenJev results look like confidence but production disagrees
OpenJev documents uncalibrated option softmax. Do not treat it as TypeSafe calibrated confidence.
Fix
- Separate environments:
prod→ TypeSafe or Gateway Jev;lab→ OpenJev. - Assert agreement on a labeled subset before any cutover.
- Keep thresholds beside question schemas in one config file.