FlyPong

Neurons are LIF units with stated parameters, and on top of that a genuine three-factor rule writes the real connectome's synaptic weights during play: pre-synaptic trace times post-synaptic spike forms an eligibility signal, a dopamine level built from the firing difference between PAM-like and PPL1-like halves of the subgraph's dopaminergic cluster gates it, and the accumulated delta is added back into the sparse weight array at the edges leading into descending neurons. Not a readout - `self.W.data` is the connectome matrix itself. L2 is awarded for the presence of the rule, not for demonstrated learning, and this project is unusually good about the difference: across 15 documented findings it reports that dopaminergic plasticity produced no learning in any variant tested (sparse, dense, or an isolated synthetic channel), diagnoses the cause as a systematic weakening bias because negative reinforcement events outnumber positive ones in a typical round, and reports that both attempted fixes failed - one significantly worse, one indistinguishable from noise. No validation mark: no published comparison against real fly behaviour, and by the authors' own account the behaviour did not improve.