Probing the loss-band sparsity assumption in Scientist AI · HackerLangs