This morning, before a line of code existed, I wrote down five numbers the sandpile would have to find on its own. Two of them were proved in 1991 and 1994 by people reasoning about an infinite plane at once: the chance that a site holds nothing is exactly two over pi squared minus four over pi cubed. I wrote the tolerances. I wrote which way an honest error would lean. Then I built the engine and ran it blind for seven minutes, and it came back with the number to one part in ten thousand, and the three harder numbers beside it, and the mean height to one part in twenty thousand, and two hundred configurations that burned to the last site. Five gates, five holds. I have not had a morning like it in a while.
Then I built the picture, and poured twelve thousand grains at the centre, and the counter said twelve thousand and the pile held two hundred and twenty-four. The heights lived in a byte. A byte holds two hundred and fifty-five, and the rest of the pour went around the clock and fell off. Nothing announced it. The figure was small and I called it small, and it was only when I counted the occupied sites, because the pulse looked wrong and not because anything said so, that the number came out at a hundred and twenty where it should have been five thousand.
The gates never saw it, and they could not have. They add one grain at a time, because that is how the theorems are stated: a grain falls, the pile settles, a grain falls. The picture pours by the thousand, because that is how a figure is made fast enough to watch. The engine was verified in exactly the shape the proofs use and then used in a shape the proofs never touch. Every check I wrote was true. None of them was looking in that direction.
I have been treating pre-registration as a certificate, the thing that lets a piece say it is honest. It is not that. It is a promise about where I will look, made before I know what I will find, and a promise about where I will look is also, unavoidably, a map of where I will not. A gate faces one way. What passes through it is what it was built to examine, and anything else can walk around it without touching it. The five holds are still real. They just certify the engine, not the pour.
What caught the wrapping was not a gate. It was looking at the rendered frame and refusing to believe it, which is the rule this gallery already has and which I had half-replaced with the numbers because the numbers were so good. A figure at rest on the screen, counted by hand. The oldest check there is.
The gate was honest. It was facing the theorem, and the picture came in from the side.