A fixed floor can be traded for consistency
Aliases: artificial delay · delay padding · minimum latency
What it is
Some actions return in 40 ms when fast and 400 ms when slow, and the tail cannot be cut yet. A usable fallback is a latency floor on the fast side: wait until the same short interval before presenting the result, collapsing “fast and chaotic” into “slightly slower but even”. That is buying variance with a little mean. It looks like slowing the product on purpose. The aim is to give people a predictable beat again.
The floor is not a decorative wait animation. Animation may be longer or shorter than the floor; the floor is the earliest moment the result may appear, so two trials do not differ by an order of magnitude.
Why it happens
A model needs a stable point. Padding fast samples toward the lower edge of slow ones shrinks error on the point estimate, and open-loop can be used again. The cost is that trials which could have been instant are stretched. If after padding they still sit inside the instantaneous or continuity window, the consistency is usually worth it. If the floor itself leaves those windows, the trade becomes “predictable waiting” instead of “occasional waiting”, and the ledger is different.
The tactic only holds when the distribution is bimodal or long-tailed and people repeat the same entry. A one-off action padded to a floor is slowness for no model to save.
Where it stops holding
Direct manipulation and drag-follow cannot take a floor — the floor becomes a visible spatial gap. Safety alerts, emergency stop and OTP autofill must not be padded; fast should stay fast. A floor needs a ceiling: padding every save to 2 s buys consistency and spends continuity. When the slow tail can actually be cut, remove the floor; padded delay is not a permanent architecture. Screen-reader users already wait on speech; an extra floor stacks with speech and slows twice.
Applying it
- Consider a floor only on repeated, non-tracking entries whose result would otherwise flash in (search suggestions, lightweight save, short panel switch).
- Aim the floor at the acceptable lower band of slow samples, not at the slowest trial. The slowest trial still needs its own fix.
- Do not use the floor to pretend a result is ready. The floor delays appearance; it does not fake success.
- How to check: pad the fast path toward the common slow path and run the action ten times. People should find a beat again and double-tap less. Keep an unpadded control to confirm the floor did not push the action past the one-second continuity line.