Raising a recognition threshold cannot stand in for an explicit engagement state

What it is

Lifting a classifier's confidence bar, or demanding a larger or longer movement, only slides along the false-fire versus miss curve. It does not turn “is the system listening?” into a state that user and system share. Engagement is explicit: the person knows they have entered a mode where commands count, and the system treats motion as input only in that mode. A threshold is a number inside the recognizer. People cannot see it, open it, or close it. Substituting threshold tuning for an engagement state quiets idle false fires and mutes real commands in the same stroke.

Why it happens

A threshold answers “does this stretch of trajectory look like an item in the vocabulary?” Engagement answers “should any trajectory count as a command right now?” The first is pattern matching; the second is session structure. Some everyday motion will graze even a high bar. Raise the bar until everyday motion cannot graze it, and deliberate moves must get larger and longer too, so fatigue and failure rise. Engagement replaces that with a discrete entry. After entry, command recognition can run at a lower bar because the candidate stream has been clutched off. Gaze selection by longer dwell is the same substitution—threshold posing as clutch—on another channel.

Studying it

Compare two systems: one that only tunes confidence or amplitude/duration gates, and one that adds an entry action separate from command gestures, then recognizes at a lower bar. Draw both false-fire versus miss curves on the same life-motion set and the same command set, and record time to complete an intentional command plus whether people can say “the system is listening.” If no operating point on the threshold-only curve meets idle false-fire and command success together, that supports “thresholds cannot replace engagement.” Count false entries separately; do not fold them into command false fires.

Where it stops holding

A vocabulary of one huge, hard-to-make-by-accident move plus a high bar can look as if engagement is unnecessary. Grow the vocabulary and the structure returns. For experts, a closed desk, or a demo where someone says “go,” the gain of engagement is eaten by the protocol. Thresholds remain useful against sensor noise; they are recognizer internals and must not carry session duties. What the entry action looks like, and how heavy it is, belongs to engagement-condition design, not to a threshold discussion.

Applying it

• Ship a listening state people can open, close, and see, before arguing about the command classifier's confidence number.
• Refuse “move the threshold from 0.7 to 0.9” as closure on Midas touch. Closure is: while not engaged, do life motions still emit commands?
• Log command events in the unengaged period, misses while engaged, and whether people can point to the current listening state.

Related

• Same group: C4.02.1 Always-on sensing makes every motion candidate input · C4.02.2 False triggering is structural, not an optimizable defect · C4.02.4 Engagement and disengagement need user-executable conditions
• Adjacent: C4.15 Conditions for entering engagement · C4.24 Recognition confidence and bias direction
• Search: engagement state · recognition threshold · explicit clutch