Mini-Batches, Vectorization, Autograd, and a Reproducible MLP Workflow

Hand-derived backward functions prove that gradients exist; a production research loop still needs three engineering contracts: mini-batches for noisy but cheap updates, vectorized tensors so those updates are matrix multiplies rather than Python loops, and autograd so you are not re-deriving every new layer by hand. This note assembles a seeded MLP workflow you can audit — the same discipline used when fitting a non-linear scoring model on a dated panel without silently peeking at the future.