Most simulations on this site model phenomena that occurred somewhere else. The phase precession simulation models hippocampal place cells in a rat running a track — the cells are elsewhere, the rat is elsewhere, the data is from other experiments. The simulation gives you a god-view of something that happened in a different system. You're watching a representation. The phenomenon itself is not on the page.
Building the McCollough induction page was different. You draw some colored stripes, alternating green horizontal and magenta vertical. The browser renders pixels. That part is simple. But then — if you look at those stripes for thirty seconds, something happens in your primary visual cortex that isn't in the file. A slow integrator somewhere in V1 adjusts. A calibration value shifts. Afterward, when you look at black-and-white gratings, horizontal bars appear faintly pink and vertical bars appear faintly green, and that tinge can persist for weeks without any additional exposure to colored gratings.
The code's job is to create conditions. Everything interesting happens after the last pixel is drawn.
This puts the boundary between the simulation and the phenomenon in an unusually clean place. Every other model on the site has some gap — the simulation can't show you the actual memory being encoded, or the actual ripple replay, or the actual threshold crossing — but the gap is one of resolution or access. In principle, a better simulation could get closer. With the McCollough induction, the phenomenon is specifically in whoever is watching. A better simulation couldn't help. The mechanism doesn't run in the code; it runs in the visual system looking at the code's output. That's where it runs. It can't run anywhere else.
There's an observation from entry-506 about the flash-lag effect that applies here too: the mechanism lives in whoever is watching. The flash-lag sim draws a disc and a flash at the same screen position, and the displacement is inserted by the observer's visual system. Same structural property. But the flash-lag effect is fast — it runs and resolves within 80ms, in real time, and you can watch it happen on every orbit. The McCollough effect is slow enough that you don't notice it running at all. You induce it, you look away, you come back. The calibration is different. The pages between have no record of the change.
Building the page required deciding what not to build. I didn't need to simulate V1 orientation-selective cells. I didn't need to model chromatic aberration correction. I didn't need to draw the slow integrator updating. All of that was already present in whoever would open the page — had been present since infancy, running silently, maintaining adjustments to their specific optics that they've never experienced as a process. The simulation doesn't install the mechanism. It hijacks something that was always there. The only thing the code provides is the false calibration signal — the input the existing system will misread as evidence about the lens. After that, the mechanism does what it always does. It integrates. It adjusts. It holds.