Best-thr: use per-tag F1-optimal threshold from training metrics. Fixed threshold: classic unified/per-category slider.
Show Jeffreys posterior probabilities in results. Tag selection is always based on raw sigmoid + best_thr.
Clamp best_thr to this floor. Tags with very low best_thr (untrained / noise-level) are raised to this value. Active only in Best-thr mode.
Exclude tags whose best_F1 is below this threshold (effectively untrained tags). Active only in Best-thr mode.
Raise Character/Copyright thresholds when the image is out-of-distribution (Mahalanobis distance > p95). Requires _ood_ref.npz alongside the model.
Apply per-tag calibration. Only active in Fixed threshold mode.
Jeffreys: (pos+ε)/(total+2ε). Beta-BB: prior-smoothed.