Faiku

The default run is a Shiu-style LIF simulation of the packed MaleCNS v1.0 graph (166,700 cells, -52 mV rest, -45 mV threshold, refractory period), and the ink-match reward drives Hige-style PAM LTD / PPL1 LTP that writes new values into the CSR weight array of the live network on KC->MBON edges, bounded to 0.05x-1.6x baseline, restricted to Kenyon cells that actually spiked. Reward and punishment are also injected as current into the real PAM and PPL1 populations, so the dopamine cells spike as well, although the gating term for the weight update is a scalar PAM/PPL1 trace rather than those cells' measured firing rate - a reviewer may reasonably want that noted. A small 8x3 motor map is trained on top as the pen, but it sits alongside the connectome update rather than in place of it. The separate --mb-only fallback is a 24-number control law with no connectome and should not be what this rating is read against. No validation mark: no comparison against real fly measurements is published, and RESEARCH.md is explicit that this is an engineered sensor/actuator sandwich around real connectome dynamics, not a fly that composes.