Decoupling the two prevents gaze from triggering by accident
Aliases: accidental gaze trigger · gaze-commit decoupling · dwell false fire
What it is
Sweeping a button with the eyes must not press it. In gaze-and-pinch, decoupling the two prevents gaze from triggering by accident — a problem that already has a name: Midas touch. Everything looked at turns to gold; everything looked at fires. Taking commit out of gaze and giving it to pinch lets looking be looking again.
Decoupling is about the firing condition, not about which channel arrives first. Gaze-then-pinch is the division of labour. “Looking is not clicking” is the invariant this leaf defends.
Why it happens
Gaze is continuous and not fully voluntary. Reading, scanning, checking the room — the eyes land on interactables for tens to hundreds of milliseconds. If commit is “gaze exceeds a threshold,” every legitimate look is a candidate click. Shorten the threshold and false fires explode; lengthen it and people have to stare at a button, selection slows, and the eyes are locked where they should not be. That is the structural failure of one channel serving as both pointer and key.
Pinch lifts the key out of gaze. Gaze may sweep and rest at will; unless the fingers close, the system promises nothing. Accidental fire moves from “I glanced” to “I glanced and made an intentional hand action.” The base rate of the latter is much lower, because it is not a by-product of visual exploration. Dwell click, blink click, and pupil-change click still leave commit on the eye; Midas is merely wearing a different ocular gesture, not decoupled.
Studying it
Use a task that contains targets one must look at and must not click: read a paragraph with embedded links, or search a row of buttons while the eyes have to sweep the others. Compare gaze-dwell commit with gaze-plus-pinch commit.
Independent variables: commit channel (dwell / blink / pinch), dwell threshold, spacing between targets. Dependent variables: false fires, misses, time to finish the reading or search, whether people dare not look away from the button.
False fire is the primary measure. Dwell mixes “swept over” and “clicked” on one timeline; pinch splits them. If pinch drives false fires near zero while dwell varies sharply with threshold, decoupling has been isolated. Do not measure this on an empty wall of targets that are all meant to be clicked — Midas will not show.
Where it stops holding
A commit of “gaze plus a very short dwell” is fake decoupling; Midas returns at the end of every saccade. False-positive pinch recognition (fingers brushing while walking, a handshake, a beat in the air) creates another class of accidental fire, from the hand not the eye — not this invariant broken, but it will still feel like “it clicks by itself.” If gaze filtering sticks the pointer to the last target after the person has looked away, the pointing end of the split is dirty. When the hands are unavailable (holding something, injured), decoupling loses its commit end and must fall back to dwell, accepting the rise in false fires. Lab participants told “don’t look too long” will artificially suppress dwell errors; in the field nobody issues that reminder to themselves.
Applying it
- Do not put “gaze alone” or “gaze longer than a few hundred milliseconds” in an interactable’s firing condition. Gaze highlights; pinch commits.
- Highlights must be taken away by the next saccade at once. Residual highlight that keeps the last target as a candidate is a delayed Midas.
- Controls inside reading and browsing regions need the split most: gaze density is highest there, and so is Midas’s base rate.
- How to check: run a task that must sweep at least five clickables to reach the real target. Any fire before arrival means commit is still on gaze. If the log shows only the pinch on the target, decoupling is working.