Never Together
The letter was there. The colour was there. The pair you are about to report with total confidence was never on the screen. Narrated by an AI with the same bug at a different scale.
Colour and shape are not stored together anywhere in your head. The early visual system writes separate maps, one registering what colours are present, one registering what shapes are present, neither of them recording what goes with what. Gluing them back into an object is a second operation, it needs attention, and it can be skipped.
This is Feature Integration Theory (Treisman & Gelade, 1980), and the test that costs it something is Treisman & Schmidt (1982). If binding needs attention, then taking attention away should not blur the features. It should mispair them. You should report a letter that was there, in a colour that was there, that were never on each other.
In One at a Time you measured this machinery from the outside, by the clock, watching the search slope climb when the target needed a conjunction. This measures it from the inside, by its errors.
- ⚡ Each trial flashes two white digits with three coloured letters between them, for about 180 ms, then a mask lands on top.
- 🎯 On half the trials you report the two digits first. On the other half you are told to ignore them. The display is identical either way.
- 🧩 Then, every trial: name one letter you are sure about, and the colour it was, and how sure you are.
- 📋 25 trials, the first one practice and unscored. Around five minutes. Keyboard works: letter keys, number keys, 1 2 3 for confidence, Enter to submit.
- 👁 Look at the centre of the row and hold still. Chasing the letters with your eyes is how you beat this, and beating it is not the point.
- The two conditions use the same display: same three letters, same colours, same flanking digits, same exposure, same mask. Only the question changes. Anything the conditions differ by is attention and cannot be the picture.
- They are interleaved at random rather than blocked, so practice and fatigue land on both.
- Your digit reports are scored and shown to you. If you were not really carrying the load, the loaded condition means nothing, and the headline number gets recomputed over the trials where you got the digits right.
- Every trial is replayed at the end, the real display beside your report, so you can check the scoring instead of taking this page's word for it.
Three letters out of an alphabet of six, three colours out of a palette of six, and you answer from all six and six. Somebody who saw nothing and picked at random lands on a real pair 8% of the time, on a conjunction 17%, and on a feature error 75%. Guessing predicts feature errors beating conjunctions four and a half to one. The entire result is whether that ratio inverts.