E.1 Reproduction and error measurement🔗

All experiments use seed 20260908. Errors are supremum errors over 100 or 200 test points, averaged over independent trials. Least-squares fits of log error against log width use N=4,8,\ldots,4096. The dashed reference lines show N^{-1/2}. These are finite test-set errors; they do not certify the supremum over a compact set. Within each trial the same sequence of neurons is accumulated as width increases. The test set is fixed within each experiment and resolution. The supplementary experiments.py script reproduces the reported results with the fixed seed.