Connectome Fighter

Neuron dynamics are a pinned Shiu et al. LIF with explicit v_th / v_rst / tau / refractory / delay parameters over ~156k MaleCNS v1.0 neurons, which earns L1. Not L2, on three independent grounds. First, no synapse changes during play: run_game_malecns_lif.py says so in its own CLI help ('no weights change during play'), and the update script calls itself a post-match updater that replays an already-logged spike file. The learned multipliers reach the simulation only by being baked into a per-character connectivity parquet beforehand, which is offline fitting followed by inference. Second, the in-memory Brian2 write that the strongest evidence points at, apply_character_plasticity, is never called from anywhere in the repository. Third, the modulatory term is a terminal game reward restricted to -1/0/+1, and the module's own docstring states that using a single game reward as the modulatory signal is an engineering abstraction - not a modelled neuromodulator. The design is careful and honest: topology and transmitter sign are held immutable, the plastic edges are genuine KC->MBON candidates from the connectome's class metadata, and the rule is anti-Hebbian in the direction fly mushroom-body plasticity runs. The README itself reports the arm as off ('continuous production reward-driven learning: OFF'). A second, better-grounded arm exists in valence_plasticity.py, cited to Aso et al. 2014 and wired into the continuous-training workflow, but it is post-match in the same way. Not validated: no published comparison against real fly measurements.