webgpu-fly

Two real connectomes (FlyWire brain, 139k neurons; MANC nerve cord, 23k) run as leaky integrate-and-fire networks in a WebGPU compute kernel, with a two-state alpha synapse and constants taken from Shiu et al. 2024 (20 ms membrane, 5 ms synapse, -52 mV rest/reset, -45 mV threshold, 2.2 ms refractory). Membrane voltage is integrated, thresholded and reset each step in `lif.wgsl`, so L1. Not L2: the weight array is bound to the kernel as `var<storage, read>` - read-only at the hardware level - and neurotransmitter sign is baked into the weights at build time, with dopamine, serotonin and octopamine mapped to zero, so there is no neuromodulation to gate a rule with even in principle. The one optimisation loop in the repo, the 'ARS/CMA-ES evolver', searches an 8-dimensional hand-written tripod-gait policy (frequency, joint amplitudes, swing ratio, stride gain, drag) and the bundled RL walking policy is an imported pretrained MLP - both live entirely outside the connectome, which is the classic trained-readout pattern. No validation mark: the project's own LIMITATIONS.md says the single dynamics check is Kenyon-cell sparsity in the 5-15% band and that w_syn was tuned to land there, making it a calibration target rather than an independent validation against fly measurements.