Record dataset/version, preprocessing, seed, model configuration, training procedure, metrics, and evaluation split.
