mindmeld-with-fly

The MaleCNS graph drives a sparse leaky-tanh reservoir: x <- (1-alpha)*x + alpha*tanh(Wx + drive) with a stated leak of alpha=0.15, so each node carries a state that decays and integrates across ticks. Under the scale that is a simplified leaky rate unit with stated parameters, which earns L1 - and the README is unusually honest that 'reservoir dynamics are not biophysics' and that the fly 'EEG' bands are spectral content of the reservoir, not electrophysiology. Not L2: W is built once from MaleCNS synapse counts (log1p) signed by neurotransmitter identity, then globally rescaled to spectral radius 0.9 at graph-build time, and is never written again - there is no optimizer, no `loss.backward()`, no learning rate and, unusually for a reservoir project, not even a trained linear readout. The EEG coupling is a fixed seeded input matrix Win that injects band power as drive; it changes what the reservoir is fed, not what it is wired with.