Learnability, recognition reliability, and bodily comfort are three separate kinds of evidence

What it is

Evaluating a mid-air gesture means collecting three kinds of evidence separately. Learnability (including guessability): after no teaching, or after one lesson, can people produce the agreed movement. Recognition reliability: how stably the sensor and model separate that movement from others and from non-input, in the target population and setting. Bodily comfort: whether muscle, joints, and posture can last for the real task duration. They answer different questions. A single “easy to use” score on a questionnaire replaces none of them.

Why it happens

Learnability lives in memory and metaphor: whether action and function line up, whether the move collides with a cultural gesture. Recognition reliability lives in sensing and classification: lighting, occlusion, individual kinematics, shape distance inside the vocabulary. Comfort lives in biomechanics: moment arms, isometric contraction, repetition count. A gesture can be guessed immediately, confused with another under side light, and still demand an overhead hold. Conversely, a hard-to-remember, practice-needed move can be distinctive enough to recognize stably and short enough to be relatively cheap. Folding three causal stories into one “gesture quality” number leaves later work unable to tell whether to change the lesson, the model, or the posture.

Studying it

Measure three times. Learnability uses elicitation or guessing: give the function, let people move first, compare with the specified form, or test retention after one lesson. Recognition reliability uses a continuous stream in the target environment, reporting confusion, segmentation error, and idle false fires, stratified by people and lighting. Comfort uses ache, endurance metrics, or EMG over task duration, not “are you tired?” asked right after one success. When the same people do all three, report a 3-D scatter, not a weighted mean. A white-wall lab, young adults, three samples each, inflates the last two dimensions together.

Where it stops holding

Expert learnability is flattened by practice and cannot argue guessability for a public setting. Recognition reliability moves with device and firmware; paper numbers are not a product gate. Comfort flips with support, sitting, and short tasks. An accessibility constraint can make one dimension a hard veto: if the shape cannot be formed, the other two do not matter. All three passing still allows refusal in public; that is social acceptability, which these three do not cover.

Applying it

• Keep three columns per vocabulary item: guessable/learnable, in-situ recognition, comfort over task duration. Missing a column bars the item from the ship list.
• When something fails, pull only the lever for that column: teaching or metaphor, sensing or vocabulary spacing, posture or support. Do not answer an ache complaint with “overall recognition tuning.”
• Ban demo recognition rate as the only review artifact. Attach at least one no-lesson guessing pass and a fatigue log from a minutes-long task.

Related

• Same group: C4.05.2 The three cannot be merged into one “ease of use” score · C4.05.3 Learnable is not reliable: intuitive moves can still be misclassified · C4.05.4 Low effort is not comfort: small moves can still load posture
• Adjacent: C4.12 Fatigue cost of mid-air gestures · C4.14 Gesture vocabulary size limits
• Search: gesture learnability · recognition reliability · bodily comfort