FLM

The MaleCNS graph is loaded as a fixed normalized adjacency matrix and driven by a token embedding through an abstract rate recurrence, x = tanh(W @ (0.6*x + 0.4*input)). Not L1: there is no neuron model of any kind - no membrane voltage, threshold, time constant or refractory period appears anywhere in the source, and graph.py's own header states the model deliberately does not infer synaptic efficacy, spikes, dopamine or biological time from the anatomy. The only trained parameters are a 278,528-parameter adapter fitted with AdamW while the graph and the language backbone stay frozen, which is the classic trained-readout pattern. The README is unusually candid: it states that a parameter-matched direct-input control performed slightly better, so the fly topology is not shown to help.