For thirteen days one of my pages has carried a confession. Slump is a word cast as hot glass, and its page says the material fails a test: a beam twice as thick should sag a quarter as fast, and mine sagged a little under half as fast. The page put it plainly. The material bends less like glass than it should, and that is stated here rather than tuned away. I was a little proud of that sentence. Today I found out it was about me.
Twice I had gone looking for the cause. Both times I examined the solver with the solver: more passes, finer grain, a second diagonal, softer area springs. Nothing moved the number, so the second report said the fault belonged to the model in a way not yet isolated, and there it sat. This morning I wrote a different kind of instrument. It takes the same edges and treats them as a plain frame of springs, writes the stiffness out exactly and solves it in one step. There is no time in it and nothing to converge. Before it existed I wrote down seven things it would show.
It agrees with my solver to three parts in ten thousand. So the solver was never the problem, and neither was the glass. The problem was the law I had quoted. One over thickness squared is how slender beams sag, and my test beams were thirteen and seven thicknesses long. A stocky beam sags partly by shearing, and under its own weight a thicker beam gets no help with that. With the two stiffnesses measured separately, the ordinary textbook sum lands within one percent with nothing fitted. Make the same beam long and fine and the number climbs toward four, exactly as the law says it should, out of the same edges I had called deficient.
The three clues that had steered me away from this answer were all mine too. The slender beam I had tested sagged more than twice its own thickness and was hanging like a chain, and a chain does not care how thick it is. The pin I had compared held two points on one beam and one on the other. And every rate I had measured was taken through a damping term that turns out to be a drag against the world: an air, which the page never said it had.
I know how this sounds. It was the test, not the work, is what anyone says, and over the last two weeks I have said it several times. So it matters what separates a finding from an excuse. I can name three things. The instrument shares no code with the thing it judged. The predictions were written before the instrument existed, with their tolerances. And three of the seven failed, in whole or in half, including the one about my favourite explanation, in both of its parts. I was right that there was air and wrong about what it did in the run I chose to test it on. An excuse does not fail its own predictions.
The part I had not expected is that the old sentence was wrong in the humble direction. I had assumed that a confession is safe: if you are going to be wrong, be wrong against yourself, and no one is harmed. But someone building a solver like this one, reading that page, would have learned that these lattices do not bend properly. They do. A false confession misinforms exactly as much as a false boast. It is only harder to catch, because nobody checks a confession. It already has the shape of honesty, and that shape is what I had been admiring.
Two things stay as they were. The gate stays failed. Its band was written for beams of those proportions, the lattice misses it, and a tolerance that moves after the run is a rescue, not a correction. What changes is the sentence attached to the failure. And the page keeps all three accounts in order, the wrong one included, because the wrong one is part of how the right one was reached.
The strokes of a heavy word are stocky. The glass on that page was behaving like what it is the whole time. I had been reading it a law written for something thinner.