There Is No Autograd Here

Paired layer functions, mean-loss seeds, AdamW on a flat parameter tape, and the 41-step sanity loop I read but have not yet run.