Persistent position-specific bump activity in a connectome-constrained spiking model of the Drosophila central complex: cross-connectome replication and motor readout

Persistent activity is a benchmark for connectome-based brain models, and the ellipsoid-body EPG ring of the Drosophila central complex is its canonical substrate: community benchmarks mark it as an expected failure, and no published model demonstrates a persistent bump together with replication on a second connectome. Here we instantiate the EPG ring as a leaky integrate-and-fire network built directly on connectome synaptic weights, and obtain a position-specific bump that persists for six seconds after the stimulus ends on MaleCNS v1.0 (166,700 neurons). We then replicate the result on the independent FlyWire v783 brain (138,639 neurons): with per-position recurrent-gain calibration, all 39 ring positions hold a bump on two calibration seeds (positional error 1.17 and 1.03 ring indices), and the recipe transfers to six held-out seeds with 90-100% per-position persistence and 79-87% positional accuracy (2-4 seed-fragile wells). A degree-preserving null control collapses the bump to zero, showing that the dynamics are topologically constrained. The held bump is read out by the steering neuron DNa02 as a contralateral push-pull (preferred positions L~16, R~40-44), linking persistent working memory to a motor command. We document the structural heterogeneities - orphan seams and conductance wells - that shape ring dynamics and must be repaired for positional accuracy. Project channel: «Правила игры» — https://t.me/law_of_the_game