Fruit Fly Plays Fruit Ninja

A 4,386-neuron induced MaleCNS subgraph used as fixed wiring for a signed rate model: six iterations of h = tanh(1.6 * W h + drive) per observation, with incoming-normalised measured synapse counts in W. L0 rather than L1, because no neuron is simulated over time: h is re-zeroed at the start of every observation, and the repository contains no membrane variable, no leak, no time constant, no spike threshold and no reset - a grep across src/ and scripts/ for membrane, tau, refractory, threshold and spike returns only math.tau in the 3D scene code. The authors state it themselves: 'Six tanh propagation steps start from zero on each observation. There is no persistent neural memory, explicit spike timing, or biological latency claim.' Not L2 either: W is loaded from a .npz and only ever read, and the only fitted component is a ridge readout over descending activity trained offline on generated scenes, which the project labels in both the training script and its summary as 'fixed biological edges; no synaptic plasticity or reinforcement learning'. Documentation quality is unusually high - explicit rewired, cut, blinded and pixel-heuristic controls with held-out seeds, and the project reports its own negative result that the real wiring did not meaningfully outperform the rewired control.