CyberFly

The full 166,700-neuron MaleCNS is simulated as the Shiu et al. 2024 LIF model, on a Metal GPU kernel in this repo whose state is membrane voltage, synaptic conductance, adaptation and refractory counter, validated bit-identical against the CPU kernel. Not L2, and confirmed at the call site rather than taken from the README: every step is invoked with `learning=False`, the engine's own docstring says 'Weights must stay frozen: the connectome is shared, a learning fly would write it', the Metal kernel raises outright if weight writes are requested, and ADR 02 records the decision as 'Plasticity is off.' This project is also a clean instance of the dopamine trap: its 'reward' channel is an injected current into the real PAM11 cells and 'punish' into PPL101 — labelled 'artificial dopamine (disclose on the board)' — which is stimulation of dopamine neurons, not a synaptic learning rule. Note that the LIF kernel's reference implementation comes from the fly-wirehead package cloned at a pinned commit into the runtime folder, so the canonical neuron model lives upstream; the Metal reimplementation and its constants are in this repo.