A single first-layer receptive field, trained by nothing but blob proposals under Boltzmann acceptance (the objective here is a fixed digit template, standing in for "classification improved"). The move: pick a center, multiply the weights under a Gaussian blob by (1 + a·G), renormalize the row — the simplex is never left. Early, wide blobs at high temperature rough in mass and accept freely; late, sub-pixel blobs at low temperature sharpen strokes and mostly get rejected. Coarse-to-fine, exploration-to-exploitation, one schedule.