FLYcasso

The full MaleCNS v1.0 graph (166,700 neurons, 25.6M connections, signs from GABA/glutamate) carries a per-neuron state that evolves under a documented leaky-tanh update with a stated leak constant, iterated three times per denoising step, and in the final configuration the direct input path around the synapses is removed so category and timestep information must cross the fly connections to reach the image or stroke decoder. Not L2: the graph is a registered PyTorch buffer, not a Parameter, propagated through a custom autograd Function literally named FixedSparseMultiply that returns None for the weight gradient - the connectome is frozen by construction. What is learned, by Adam and backprop offline, is per-neuron gain and bias plus the U-Net adapters and decoders; per-neuron gain is intrinsic excitability, not a synapse, and an optimiser's gradient is not a fly learning rule. The model card says this plainly in its own words. This is a borderline L1 worth a second reader: the state update is documented and the state genuinely evolves, but three iterations inside one forward pass of a diffusion model is a recurrent layer whose sparsity pattern is a connectome rather than a simulation of a fly over time - there is no membrane potential in mV, no refractory period and no synaptic delay. The model card is also admirably explicit that earlier checkpoints without the synaptic output path barely changed denoising error, i.e. the connectome was doing little.