False triggering is a structural property of the modality, not a defect to optimize away

What it is

A false mid-air fire is not leftover noise from an under-trained model. It is a structural property of the modality: while the sensor is looking, the person is moving, and vocabulary shapes also occur in ordinary life, a false event has positive probability. A larger network, more data, or a higher confidence threshold can suppress one kind of wave. It cannot suppress the fact that motion is both living and commanding. Treating the problem as a defect sends a product into endless tuning without ever adding a clutch that splits aiming from firing.

Why it happens

The distribution of everyday motion covers the gesture vocabulary. A nod, a goodbye wave, reaching for a cup, holding up “three,” can all land inside a recognizer's decision region. Classifiers learn boundaries on shape and trajectory, not intent. More capacity fits the training distribution more tightly; it cannot stamp motion with a “this time it counts” label, because that label does not exist in physics. False fires and misses trade off: exclude the casual goodbye wave and the deliberate “next page” wave gets harder too. So false triggering returns as the vocabulary grows and the setting gets more lifelike. It is not a bug a version number can shut forever.

Studying it

Treat false fires as a primary measure, not an appendix next to recognition rate. Collect a “life motion” set: conversation, walking, tidying a desk, giving directions, run the recognizer without briefing the vocabulary, and get false events per minute. Collect deliberate gestures from the same people and get misses. Draw the false-fire versus miss curve, not a single operating point. After a new model or more data, remeasure the curve. If the curve barely translates and only the operating point slides, that supports structure rather than underfitting. Field counts should also be stratified by consequence: skipping a track and killing a light must not be summed into one number.

Where it stops holding

A tiny vocabulary of highly artificial moves (fist then a star nobody draws in a kitchen) can drive false fires down to a shippable level. That trades learnability and fatigue; the structure did not vanish. A closed desk, one person, a still background, makes the curve look good until it moves to a living room. Blaming every false fire on “the user being non-standard” refuses to admit that everyday motion is legitimate. Gaze false selection is also structural, but the trigger is dwell time; importing that dwell threshold does not mean the mid-air problem is solved.

Applying it

• Write an acceptable idle false-fire rate in the requirements (how many per hour, at which consequence) and treat it as a ship gate, not a post-launch polish item.
• Refuse “train another version and false fires will be eradicated” as the only countermeasure. The plan must include a clutch, an engagement state, or moving high-consequence actions off the mid-air path.
• Every model change must run both the life-motion set and the deliberate set, and publish the false-fire versus miss curve. Do not show only accuracy on deliberate gestures.

Related

• Same group: C4.02.1 Always-on sensing makes every motion candidate input · C4.02.3 Raising thresholds cannot replace explicit engagement · C4.02.4 Engagement and disengagement need user-executable conditions
• Adjacent: C4.24 Recognition confidence and bias direction · C8.03 Midas touch in eye input
• Search: false positive gesture · structural Midas touch · intention versus motion