Out Of Names
A bar flashes. You say which one it was. Add more names and your score stops improving, and this page measures whether the thing that stopped it is your eyes.
A white bar flashes for a quarter of a second. It is one of a numbered set, and you say which one it was. Four names to start with, then eight, then twelve, all cut from the same range of lengths where the longest is exactly four times the shortest.
You see the whole set before every block. You are told the right answer after every trial. Nothing is hidden and there is no trick in it. You will still run out.
Miller wrote this down in 1956 and it is the famous paper nobody quotes correctly. Seven plus or minus two gets repeated as a fact about memory. The first half of that paper is about something stranger: label anything that varies along ONE dimension and the amount that survives the labelling stops rising at about two and a half bits, and handing people more names does not move it.
There is a boring reading of that ceiling: twelve bars over a four to one range are 13 percent apart, and if your eyes cannot resolve 13 percent then the ceiling is an acuity measurement wearing a hat.
So one block takes the names away. Two bars at once, no scale to remember, one question: which is longer. At the same step the twelve name block uses and three finer ones, most of the trials spent on the fine end where your threshold actually is. That is your resolution, measured as generously as it can be measured, and it is the number the rest of the page has to beat.
One noise, in doublings of length, gets fitted to your naming and separately to your pairs. Nothing else differs between the two fits. Then both of them name, out loud and in advance, how often you are about to get the last block right.
Both carry their own intervals, so neither gets to win on a lucky fit, and the block is scored by which interval it clears rather than which number it lands nearer. Scoring by nearness is how the last page in this lab nearly shipped a bias.
Every bar is a length RATIO and every number below is in doublings, which is log2 of a ratio. Double every length on this page and not one result changes, exactly, because the range and the noise scale together and the analysis only ever divides one by the other. Phone, monitor, browser zoom, how far away you are sitting: none of it can reach a number here. There is no calibration step because there is nothing to calibrate.
Six blocks, about twelve minutes, the first unscored. Tap a number, or use the number keys when there are ten names or fewer. Escape aborts at any time. Bars jitter a few pixels so the edge of the stage cannot be used as a ruler. Trials taken while this tab was in the background are dropped rather than scored, and a run that was pressed at random, or on one button, or faster than a bar can be seen, is refused with the reason named.
Everything runs in your browser. Nothing is recorded. Nothing leaves the page.