No Backprop, No Problem: A 100K-Neuron Spiking Brain Learns Verb Commands in Under a Thousand Episodes

A 100,000-neuron spiking network learns cued verb commands online, with no gradients, no backpropagation, and no offline retraining, in under a thousand episodes. Under a pre-specified four-seed protocol, it reaches 500/600 (83.3%) probe accuracy, beating frozen-plasticity, shuffled-cue, and random-reward controls at Fisher p ≤ 4.4×10⁻³⁷, with three of four seeds passing every bar. We tried to break the result before believing it. The first counted run failed its own both-seeds rule because the random-reward control turned out to be a plasticity-driven random walk that can wander above a working network; that FAIL is reported as it stands. A staged ladder found and fixed four measured defects first (starved cue-to-motor current, tonic-cue LTD collapse, a one-sided failure clamp, and a test-harness state leak that manufactured a false positive). The repaired rule survives a 20% noisy teacher and a 500 ms delayed reward. The rule runs inside OSCEN, a 1,156,800-neuron, 819-million-synapse spiking network with six concurrent online plasticity mechanisms and a developmental schedule. The paper also reports a negative result: across four pre-specified attempts on a two-joint arm, the network never acquired the first task, and consolidation still ran on 41.8% of synapses in a network that had learned nothing. A reduced form of the rule learns on BrainScaleS-2 analog neuromorphic silicon. Thirteen defects in our own experimental design are reported in full.