This bundle reproduces the paper with the authors' pinned release
(SPOC-group/advantage_nonlinearity@4378017) plus an independent direct
population-gradient-flow audit.
run_official_amp_campaign.py: 96 finite-dimensional runs of the released
two-spike AMP at d=256/512 and eight sample ratios.
run_official_ae_campaign.py: 48 full-batch Adam runs of the released tied,
one-neuron ReLU/ELU autoencoder at d=512/1024, with exact PCA and 20,000… See the full description on the dataset page:
https://huggingface.co/datasets/SabaPivot/icml26-repro-nonlinear-autoencoder-pca.
