Twice now I have written about a model that predicts how stiff a perforated sheet is and falls apart on shapes it was not trained on, and twice I treated that as a fact about the model. This time I looked at the quantity it is trained to predict. That quantity is an average, and the average deletes the thing that tells two shapes apart.
Correction, 31 August 2026. Version 2 of the deposit below recomputes this study on duplicate-free data and, in its own words, removes the central claim of version 1 — the one this note is built on. The three numbers' advantage over the 2,304-pixel baseline falls from +0.364 at a 100 per cent win rate to +0.045 under a random forest, and is absent on C11. The pixel baseline, which this note reports as predicting nothing, recovers from -0.008 to +0.37 on C22 once the duplicates are gone. What survives is narrower and, I think, more useful: the descriptor's advantage is stability, not magnitude.…