FLYT3

Two brains, and they sit on different sides of the L1/L2 line. The whole-brain arm is a batched GPU LIF over the MaleCNS graph (166,700 neurons, 25.6M retained connections; leak 0.85, gain 0.15, per-neuron homeostatic threshold with target rate 3% and gain 12, reset to zero on spike) whose weight tensor is built once and only ever read under torch.no_grad, with a REINFORCE-trained two-layer policy head over motor rates and a 'punish' method that merely adds an aversive current to the PPL101 pair - that arm alone is L1, and it is the one the web app serves. L2 is earned by the second arm, flyc4/mushroom.py: the real MaleCNS KC->MBON synapse-count matrix is the learned tensor, MBON compartments are labelled approach/avoid from the connectome's own PAM and PPL1 innervation following Aso et al. 2014, and dopamine_update writes dw = -eta * KC * signed_dopamine clamped to [0, w0] under no_grad - a coincidence rule between an active Kenyon cell and a compartment-routed dopamine signal, citing Handler et al. 2019 and Bennett et al. 2021, with no optimiser anywhere near it. It runs during self-play, once per finished game, and it is shipped rather than dormant: deploy/mb-train-job.yaml is its own training Job and MODEL_CARD.md reports mb_weights.npz at 1,000,000 games and 74% win vs random. Two caveats a reader should have: the README predates this arm and still claims the wiring is never modified, and calibrate() rescales the connectome counts per MBON to normalise drive before learning starts. Not validated - the model card says the dynamics and dopamine gating are engineering choices, not validated biology, and the honest fire-detection note stands too, where a pooled-pixel MLP with no circuit beats the circuit decoder 85.7% to 60.8%.