Concepts

The dichotomy between objective and perceived performance, the 20 % just-noticeable difference, and where looking-fast stops being a substitute for being-fast and becomes a tax on the user's trust.

Active vs. passive phases, the dopamine pathway, filled vs. empty duration, prospective vs. retrospective duration judgments, and the one-second active-to-passive transition.

Miller 1968's 17-transaction taxonomy, Card et al. 1991's perceptual / immediate-response / unit-task tiers, Doherty's 400 ms productivity cliff, and Nielsen's 1993 distillation. Why the clean 0.1 / 1 / 10 trichotomy is Nielsen's framing, not Miller's.

Decompose every wait into pre-action signal → response → animation → completion. What you can tune at each stage. The tip-the-hand rule.

Anstis on contrast and motion. Harrison et al. on backwards-decelerating progress bars. The geometric-mean indifference threshold. When illusions are honest and when they cross into deception.

Eizenberg's argument, latency JNDs in direct manipulation, and where polished placeholders become polished lies.

INP and Web Vitals as a starting point. How to add perception to a quantitative budget. The role of adaptive loading and the performance scaler.

The decision rule: nothing under 1 s, spinner 1–2 s, skeleton or determinate progress 2–10 s, engagement past 10 s. Spinners rehabilitated, the 500 ms delay gate, the prospective–retrospective trade in engaging loading, and empty-state hygiene.

AI waits are not the deterministic page-load waits the rest of the platform is tuned for. Three properties change: duration variance across two orders of magnitude, conversational shape, and the answer arriving mid-wait. Which classical thresholds survive the move to AI, and which do not.

Where AI perception engineering crosses into deception. Fake streaming, manipulated cadence, confident UX over uncertain output, hidden tool calls, cancellation theatre. The line between polish and lying — and how to stay on the right side of it.