Re-prompt intervals should lengthen each time
Aliases: prompt backoff · increasing wait · stretched re-prompt gap
What it is
A smart display asks who the recipient is, waits five seconds, asks again, five more, five more. Someone still flipping through contacts is sliced by the beat. Someone who already left has questions inserted into the house on that beat. A lengthening re-prompt interval requires the second wait to outlast the first, and the third to outlast the second — not the same metronome. It changes the spacing of the clock. It does not change how the question is worded, and it does not replace “after how many asks we stop.”
Why it happens
After the first no-speech, two facts have changed. If the person is still composing, the original time budget is already spent; the next budget needs to be larger, not equal. If they have left, speaking on the original cadence raises the social cost in an empty room and almost never buys the task back. Lengthening the gap is backoff: more time for “still thinking,” a lower rate of floor-taking if “already gone” is true.
An equal beat sounds like nagging because it treats every blank as the same failed retry. Backoff admits that a later blank already contains the information that the earlier ask drew no speech. A longer gap is not courtesy decoration. It moves probability mass from “ask again now” to “wait a bit more.” A stop rule is still required. Backoff without a stop decays into sparse waiting that never ends.
Studying it
Lock wording and the maximum number of chases. Compare constant interval with increasing interval (for example 1×, 1.6×, 2.5× the class timeout). Dependent measures: late onsets cut by the second prompt, how many prompts an empty-room camera hears, annoyance ratings. Do not rewrite the second question in the same arm, or spacing cannot be separated from new wording.
In the field, measure wall-clock gaps between adjacent system questions. If the second-to-third gap is smaller than or equal to the first-to-second, backoff did not happen. Exclude long system-side silence while a database is queried.
Where it stops holding
Evacuation, medical alerts, safety actions already counting down should stay dense and even, not back off. Some users use a regular metronome as a memory scaffold; they should be able to switch back to equal gaps rather than losing re-prompts. Backoff is not a substitute for ending: a fifth ask after thirty seconds still occupies the floor. On very short yes/no items the first wait is already tight; stretching further becomes fake dead air, so the growth ratio should be gentler.
Applying it
- Write the wait sequence as increasing. The second gap must exceed the first; the third must exceed the second. Never shorter than the first.
- Multiply the prompt-class baseline. Do not invent a second set of seconds that ignores class.
- Store timestamps for each chase. Assert in the same session and slot that Δt is monotonically non-decreasing.
- Find sessions whose onset falls just after the first timeout. If those utterances complete under backoff and get sliced under equal interval, the spacing policy is doing its job.