Flyhard

Neurons are sparse recurrent rate units with an explicit bounded leak, state = (1 - leak) * state + leak * tanh(signal), run over the immutable measured MaleCNS adjacency - documented dynamics, so L1. This project is an interesting inversion of the usual trap: the readout decoder is a frozen random projection that is explicitly untrainable, and what actually gets trained is a per-edge gain on the connectome's own measured edges. It still does not reach L2, for two reasons. First, the rule is supervised regression - Adam on a mean-squared error against target joint angles - not anything grounded in fly biology; a repo-wide search for plasticity, dopamine, STDP, Hebbian or eligibility in src/ and scripts/ returns nothing. Second, it is offline: training writes a checkpoint, and the CARLA run loads that checkpoint under .eval() and torch.no_grad(), so no synapse changes while the fly is driving. This is the classic bake-in-fixed-weights near-miss, just applied to the connectome instead of a readout.