Pre-registration (second protocol) of a replication of the neural scaling exponent claim of Kolmogorov-Arnold Networks with large training sets (arXiv 2404.19756)

Frozen pre-registration (2 Nov 2026) of a second private replication of the claim that KANs trained with LBFGS and grid extension reach a test-RMSE-versus-parameter-count exponent close to 4 on smooth toy functions where MLPs do clearly worse (Liu et al., arXiv 2404.19756; code KindXiaoming/pykan, commit ecde4ec3). It replaces a first protocol (doi:10.5281/zenodo.23856253) that returned Inconclusive because its training sets were too small and its MLP comparison family too sparse. Three scored functions, training sets at least ten times the largest parameter count, an analytically generated MLP family covering the KAN parameter window, a pre-registered training-success gate, an interpolation control, 20 seeds per function, pre-set bounds (KAN slope lower bound 3.0; KAN minus MLP slope above 1.0), known-answer and negative controls. Contains the decision code with its planted-effect table, harness, MLP-family generator, cost estimate and the run kernel. Deposited before any confirmatory run. Outcome wording: Replicated / Did not replicate / Inconclusive under protocol X.