Two similar mid-air shapes must not carry consequences that differ by an order of magnitude

What it is

If two items in the vocabulary sit close in sensor space, one must not be “next page” and the other “power off” or “pay.” Neighbour confusion is the recognizer's normal, not an accident. When consequences differ by an order of magnitude, one confusion is unacceptable. The fix is to pull the shapes apart, or to take the high-consequence item out of mid-air, not to hope a threshold “gets it right.” How to decouple recognition difficulty from consequence, and whether to add a second confirm, is a later layer. At vocabulary layer, first forbid making a dangerous pair as twins.

Why it happens

Off-diagonals of the confusion matrix are where shape distance was not enough. Mid-air has no keycaps to separate two commands; distance lives only in trajectory space. Production variation makes actual distance smaller than template distance. Putting a high-consequence command in a dense region raises expected loss from a false fire. Raising that item's threshold raises misses; people then go larger, closer to the dangerous neighbour. Pulling shapes apart (pinch for one, a wholly different path for the other) or handing high consequence to a key is what changes expected loss. Growing the vocabulary crowds old safety distances; neighbours must be remeasured, not assumed still safe from the original pairing.

Studying it

Measure pairwise distance and confusion, then label each item with a consequence grade. Paint red the cells that are high confusion × high consequence gap. Experimentally replace a dangerous pair with a morphologically distant substitute, and watch whether false-fire damage drops and whether learnability is sacrificed. Consequence grades must come from real harm (hard to undo, money, safety), not from a designer saying “this one matters more.” A mirror swipe pair where one browses and one deletes is a classic red cell.

Where it stops holding

Two low-consequence, reversible items (page left / page right) may tolerate neighbours. High consequence with a mandatory second confirm intercepts the neighbour risk, and the vocabulary ban can relax; the confirm itself needs its own test. User-defined vocabularies will import dangerous pairs; the system should warn on save, not forbid customization. A recognizer upgrade can change distances; old safe pairs need regression.

Applying it

• Build a table of consequence grade × pairwise confusion. Red cells must be split: reshape or leave mid-air.
• Every added item is remeasured against all high-consequence items, not only tested for its own recognition rate.
• Browsing mirror pairs may sit close. Delete, pay, and power-off may not neighbour any frequent item.

Related

• Same group: C4.14.1 The gesture set must be bounded and discoverable · C4.14.3 Gesture meaning is validated with target users, not designer intuition · C4.14.4 Learned gestures must be relearnable on demand
• Adjacent: C4.27 Decoupling gestures from consequence severity · C4.02 The Midas touch problem
• Search: gesture confusion · high-stakes command · neighbour distance