A real network, training in this page: conv 5×5·8 → pool → conv 3×3·16 → pool → dense, 3,946 parameters, on 3,000 handwritten digits at 16×16 — SGD with momentum, cross-entropy loss, nothing hidden. The curves are the four shapes the chapter's pathology gallery taught: watch for the train/test gap, and for what the learning rate does at each extreme. The eight first-layer filters redraw as they learn; they start as noise, and whether structure appears — and when — depends on the dial you are holding.