Accuracy and performance

Temporal stability and detector accuracy are different measurements. External evaluation compares raw detections, a tracker baseline, TemporalFix, and a valid tracker-plus-TemporalFix composition on identical frozen detector output.

The protocol measures box jitter, class-switch rate, gap-recovery precision and recall, false-positive persistence, detector task metrics, and added latency. No improvement is claimed unless a stored machine-readable artifact supports the exact configuration, dataset split, metric, and environment.

Smoothing trades responsiveness for stability. none has no smoothing lag; EMA is inexpensive and responds faster as ema_alpha increases; Kalman smoothing predicts constant velocity and depends on its process and measurement noise settings.