the_buzz

The FlyWire adjacency matrix becomes the recurrent reservoir of an echo state network whose state evolves under a documented leaky-tanh update with declared leak rate (0.7-0.9), spectral radius (0.9) and input scaling, which is the same class of simplified rate unit the scale credits at L1. This is the weakest form of L1: there is no membrane potential, no spiking, no refractory period, and the neurotransmitter signs are discarded entirely - edge weights are log1p-transformed, row-normalised and rescaled, leaving an all-positive reservoir. Not L2: the reservoir matrix W is built once and never written again, and the only thing fitted is the linear output vector w_out, solved in closed form by ridge regression - a trained readout on a frozen graph, exactly the pattern the scale calls out. Rate the simulation, not the backtest: the headline +4.32% versus +3.78% for the degree-preserving rewired control is a single 1.4-day test window and is not evidence of anything on this scale.