At low confidence the system must choose whether to prefer misses or false positives

What it is

When a score sits in the mid-range, the system has to pick a bias direction first: treat the observation as “does not count” (prefer a miss, false reject) or as “counts” (prefer a false accept). Direction is policy, not the operating point the trainer happened to stop on. Low confidence does not vanish; it is only pushed to one side. Tracking loss is no score to bias. This is about which way to fall when there is a score, but not a clean one.

Why it happens

Every threshold on an ROC or DET curve is a miss–false-accept trade. Mid-air gesture’s background is continuous motion and a huge negative class; the same threshold that looks “accurate” on clean lab tokens will emit a string of false fires in a living room. Bias toward misses and users retry, calling the system sluggish; bias toward false accepts and users meet effects they did not produce, calling it snatched. The two pains differ in kind and cannot be summed into a neutral “total errors.” The mid-range can also refuse to fall either way yet: delay, demand a second-channel confirm, or switch to a coarser gesture. That too is a bias—toward “not now.”

Studying it

Sweep thresholds on a stream that contains large amounts of non-gesture motion. Report miss rate and false-accept rate separately, plus subjective cost ratings for each error class. Compare a fixed threshold, delay in the mid-band, and confirm in the mid-band. The task needs deliberate background motion (talking, fixing hair); otherwise the negative class is too thin and the false-accept side cannot be measured. Plot the operating point on a DET; do not hand over a single F1.

Where it stops holding

A vocabulary of one “next” that is cheap to undo often feels better slightly biased toward false accepts than toward repeated misses. Commands that cannot be undone, or that startle bystanders, go the other way. Under extreme class imbalance, accuracy is swallowed by the majority class and cannot choose direction. If the population includes motor impairment of the hand, misses rise systematically at the same threshold; that looks like “biased toward misses” but is a body difference, and wants another channel rather than a higher threshold.

Applying it

• Write, for each command, which way low confidence falls, or that it falls to “delay.” Forbid one silent default threshold for the whole library with no stated reason.
• Do not swallow the mid-band in silence: either explicitly not-fire and allow an immediate retry, or ask for a light confirm.
• Before launch, sweep the operating point on long recordings that include background motion, and accept misses and false accepts separately rather than whether one accuracy number cleared a bar.

Related

• Same group: C4.24.1 Recognition systems report a confidence for each decision rather than a binary outcome · C4.24.3 Bias direction should follow consequence severity rather than a uniform setting · C4.24.4 Adjusting the confidence threshold is a product decision, not something to inherit from the model default
• Adjacent: C4.02 The Midas touch problem · C4.27 Decoupling gestures from consequence severity
• Search: false reject · false accept · operating point