When expanding the gesture vocabulary, re-check whether similar gestures map to different consequence levels

What it is

Every motion added to the vocabulary may sit close, morphologically, to a new or old template. If the two neighbors run “next” and “delete,” that is a consequence collision: a wobble on the recognizer’s boundary jumps between two entirely different costs. Expansion requires re-checking the whole consequence map, not only whether the new motion recognizes well on its own. Morphologically close gestures must not carry far-apart consequences—that is a vocabulary constraint. What is required here is that the constraint be rerun at the moment of expansion, because collisions grow later.

Why it happens

Template space crowds as entries accrue. The nearest neighbor of a new motion may be a high-consequence command left a year ago, when the two consequence tiers had never met. A confusion matrix is computed per class and will not cluster by consequence tier; two symmetric 4% confusions, one preview and one payment, are not two 4%s. Version iteration also likes to extend by “a similar swipe, other direction,” which is memorable for users and unstable under depth noise—exactly how consequence collisions are made. The check must tag each confusable pair with both tiers. A cross-tier pair means change the motion, add confirm, or split the vocabulary, not merely tighten the new motion’s threshold—tightening turns the collision into systematic misses on the newcomer.

Studying it

After each vocabulary change, recompute pairwise distances or confusions for the whole library and color by consequence tier. List cross-tier pairs over a distance threshold as collisions that must be handled. Inject trajectories halfway between two templates and see which tier the system falls toward. Compare how many collisions “new-motion accuracy only” catches versus “rerun the whole consequence map.” Let a designer who does not know the old vocabulary add a motion, and watch the collision rate without a check—that is evidence of a process defect.

Where it stops holding

If the new motion comes from a disjoint modality or body part (head versus finger), the shape spaces do not meet and the check can pass quickly. Personalized vocabularies (each household teaches a few motions) cannot be fully checked at the factory; run a live collision check when the user teaches a new one. Toy demo vocabularies often make motions similar on purpose, with collision as an Easter egg, and do not represent a production library. When several users share one library, A's “easy gesture” may be B's high-consequence; the check uses product-level tiers, not one person's habit.

Applying it

• Make “confusable pairs × consequence tier” a release checklist. A cross-tier pair must close before the version ships.
• When extending, do not let “the same swipe, other direction” carry cross-tier commands; leave direction differences inside one tier.
• When a user teaches a new motion that sits too close to a high-consequence template, refuse to save and say “too like delete,” rather than accept it and lean on the threshold.

Related

• Same group: C4.27.1 Recognition difficulty and consequence severity should be designed independently, not treated as naturally bound · C4.27.2 High-consequence actions should not rely on a single one-shot recognition result · C4.27.3 Low-consequence, reversible actions can accept looser recognition thresholds
• Adjacent: C4.14 Gesture vocabulary size limits · C4.24 Recognition confidence and bias direction
• Search: vocabulary growth · confusable pair · consequence collision