The estimator starts with a randomly rotated, 66,048-point Kerdock rule, followed by adaptive pruning. The paper studies the unpruned sampling core and proves a depth-32 Hermite bound leaving a 0.0233188% NNGP-risk gap among fixed positive rules with the same support budget.
Paper and reproduction materials: GitHub release for submission 327786
AIcrowd submission 327786
AI usage disclosure