Fly-Brain-AI / plastic-fly

The FlyWire connectome path is a Brian2 leaky integrate-and-fire network with fully stated parameters (v_0 and v_rst -52 mV, threshold -45 mV, membrane tau 20 ms, synaptic tau 5 ms, refractory 2.2 ms, delay 1.8 ms) following Shiu et al., so membrane voltages change under a documented neuron model. Not L2, and the 'plastic-fly' name is the thing to check here: the plasticity is real code that really writes weights online with a Hebbian-plus-homeostatic-decay rule, but it writes them in PlasticRecurrentNet, a torch RNN whose sparsity mask is torch.rand(hidden_dim, hidden_dim) > sparsity and whose weights are xavier-initialised - a random network, not the connectome. Its only caller, experiments/terrain_shift.py, loads no connectome data at all (no reference to FlyWire, Brian2 or the connectome anywhere in the file). The connectome synapses are declared as a static 'w : volt' with on_pre='g += w', assigned once from the measured weights, with no on_post and no update anywhere. So the plastic variant and the connectome variant are disjoint pipelines, and the README agrees, stating the three adaptive behaviours emerge 'from wiring alone, without learning or parameter fitting'.