fly-fpv

The MaleCNS graph is used, but only as an adjacency mask on a sparse artificial network - which edges are allowed to be non-zero. Nothing about a neuron is simulated: there is no membrane voltage, no time constant, no threshold and no spike anywhere in the source (a grep for leaky, LIF, tau_m, v_rest, membrane, refractory and spike across the Python returns nothing), and the policy carries no state between simulator steps, so each control decision is an independent feed-forward pass of residual sparse layers. That rules out L1. Edge values are not the measured synapse counts either - they are initialised from torch.randn with fan-in scaling and learned from the flight task by PPO through torch.optim.Adam, with biological transmitter signs explicitly not enforced. So it does not approach L2 either: the weights that change are an optimizer's parameters, changed offline during training, not synapses changing during an episode under a fly rule. The project states all of this plainly rather than overclaiming - 'The connectome is an architectural constraint, not a physiological simulator' - and its headline experiment is the honest one, comparing the real mask against degree-preserving shuffled, Erdos-Renyi and dense MLP controls.