Legality forces every surviving P row to renormalize back to Σ = 1, so pruning an entry doesn't just remove its contribution — it inflates every survivor by 1/m (m = the mass that was kept). Panel A: drag the slider to one surviving edge and that synapse becomes a full-strength wire; on the fixed input shown, the neuron's activation jumps far above anything downstream weights were trained to receive. Panel B is why the textbook saliency score goes below random: a global sort lets one hot row hoard the budget while the others starve down to their single forced entry and spike (rust bars = renorm factor). Balanced allocation, same budget, keeps every factor near 1. Starved rows aren't lossy — they're toxic.