Guessable is not recognizable: a semantically obvious move can still be misread

What it is

That people immediately guess “wave means next page” says the metaphor lined up. It does not say the recognizer can separate that wave from a goodbye, a bug-swat, or pointing at someone. Learnability asks whether people can think of and produce the agreed move. Reliability asks whether the system can retrieve it stably among noise, viewpoints, and neighbours in the vocabulary. Semantically obvious moves tend to be the ones already frequent in life. The better the guess, the larger the overlap with living motion, and the higher the misread risk. Writing high agreement from an elicitation study as “so this gesture can ship” skips the recognition dimension.

Why it happens

Elicitation asks people to pair functions with movements. High-consensus candidates usually come from symbols culture already has: wave, nod, thumbs-up, swipe. Those symbols sit close together in sensor space, and close to unrelated motion. The classifier has no “semantics” channel, only trajectory and pose. Two stretches that mean opposite things to a person can be almost the same joint-angle sequence. Better teaching makes people willing to go smaller, faster, more lifelike; the template then sits farther from the real distribution, and recognition gets more brittle. Conversely, a deliberately artificial move (fist then a star) is hard to guess but far from neighbours, so recognition can be stable. Learnability pushes items toward the centre of living motion. Reliability wants them on empty ground. The directions oppose.

Studying it

Run guess/elicitation first for agreement or guess rate; then drop the same candidate into continuous recognition for confusion with neighbours and with living motion. Plot guess rate against confusion; high-guess, high-confusion points are the failure this dimension exists to catch. A control: slightly reshape a high-consensus move (add a prefix pose rare in life) and watch how much guess rate drops versus how much confusion drops. Isolated-clip classification accuracy hides the living overlap. Report “people guessed the function” and “the system retrieved the instance” as different facts.

Where it stops holding

In a closed, taught setting with two or three items, teaching can sit on the confusion of an obvious move. In public, untaught, with a larger vocabulary, the structure returns. User-defined vocabularies raise learnability and push conflict onto the user; neighbour problems on the system side do not vanish. Gaze or speech used as a clutch cuts the living overlap of an obvious hand move; that is multimodal, not the gesture becoming reliable on its own. A move the designer finds obvious that means something else in the target culture hits learnability and social life together, not only recognition.

Applying it

• Passing elicitation or guessing only enters a candidate list, not a ship list. Before shipping, require confusion data against living motion.
• For high-consensus candidates, add an observable prefix or clutch that is not part of the command, rather than hoping the model “learns to tell goodbye waves from next-page waves.”
• Do not write “natural, therefore accurate” in product copy. Natural means guessable. Accurate needs other evidence.

Related

• Same group: C4.05.1 Learnability, recognition, and comfort are separate evidence · C4.05.2 The three cannot be merged into one “ease of use” score · C4.05.4 Low effort is not comfort: small moves can still load posture
• Adjacent: C4.02 The Midas touch problem · C4.14 Gesture vocabulary size limits
• Search: guessability · gesture confusion · elicitation agreement