Bias direction should follow consequence severity rather than a uniform setting

What it is

“Next slide” and “submit payment” may share a recognizer. They must not share the way low confidence falls. Bias direction follows consequence: low-consequence, reversible commands can lean slightly toward false accepts for flow; high-consequence, hard-to-undo commands should lean toward misses, or demand confirm in the mid-band. One threshold for the library hides consequence differences inside a model default. What is being changed is how scores are cut, not how hard the gesture is to perform—a hard gesture is not a substitute for a stricter threshold.

Why it happens

The same ambiguous trajectory, cut as “page” is one extra flip; cut as “delete” is an irreversible commit. Expected loss is asymmetric in consequence: missing a page costs another wave; a false delete costs a recovery flow. A uniform threshold assumes both error costs are equal for every command. Setting the operating point per command, or switching it with UI state (loose while browsing, strict on a dialog), aligns loss with product intent. The switch has to be predictable: users must not find that the same pose suddenly went from “always works” to “never works” because a confirm dialog happened to open.

Studying it

Bind the same motions to a low-consequence and a high-consequence command, sweep thresholds, and record misses, false accepts, and the subjective unacceptability point separately. Then compare “one threshold for all” against “tiered by consequence” on time and error cost. When UI state changes (canvas to delete-confirm), measure whether users understand that sensitivity changed. Do not use gesture completion time as a proxy for consequence—fast is not low-risk.

Where it stops holding

When every command is one-tap undoable and undo is obvious, tiering pays less and a uniformly slightly loose threshold may be easier to learn. When every command is high-hazard (an industrial stop), there is no loose side to tier toward—only miss-biased plus a second channel. Automatically tagging “looks hard to do” as high-consequence smuggles gesture difficulty in as severity, and the tiers tangle with vocabulary design. Children and adults price the same consequence differently; adult-rated tiers do not transfer to a family device.

Applying it

• Label a consequence tier on each gesture command and write that tier’s default bias (loose / mid / strict / mid-band confirm). Do not let the recognizer’s single threshold override this table.
• When the UI moves from a low-consequence region into a high-consequence one, show a visible confirm or a sensitivity-change cue, so the same pose does not suddenly feel broken.
• At acceptance, pick at least one low- and one high-consequence command and replay the same ambiguous trajectories: the former may false-fire a little; the latter must not fire at the same score.

Related

• Same group: C4.24.1 Recognition systems report a confidence for each decision rather than a binary outcome · C4.24.2 At low confidence the system must choose whether to prefer misses or false positives · C4.24.4 Adjusting the confidence threshold is a product decision, not something to inherit from the model default
• Adjacent: C4.27 Decoupling gestures from consequence severity · C4.14 Gesture vocabulary size limits
• Search: consequence · per-command threshold · risk-sensitive bias