Which build of the connectome a whole-brain fly model runs on changes what it can do

Whole-brain spiking models of Drosophila are built from "the FlyWire connectome", but that is not one graph: synapse tables are re-predicted, edges thresholded, and each neuron's transmitter, which fixes the sign of everything it does, is inferred, and left blank where the inference is weak. We built flybench, a benchmark of 37 tasks drawn from published fly experiments, scored against shuffled-wiring controls, with predictions committed before the first run for every task and model change after the first 15. We used it to compare three builds of the same 139,255 neurons under the widely used leaky integrate-and-fire (LIF) model of Shiu et al. (2024). At the gain where Shiu et al.'s graph passes its core reflexes, the build changes the mean task score by as much as adding spike-frequency adaptation to the neuron model does (+0.098, 95% CI +0.042 to +0.164, against +0.092); at the reference model's own gain the builds do not differ detectably. Signing the 19,658 neurons whose transmitter the public release leaves blank, with every edge unchanged, scores about as high as Shiu et al.'s graph. It restores part of the antennal lobe's channel identity, and at gains 0.5 to 0.65 it is the only change under which a model steers away from CO2 while still steering toward attractive odours. It also moves the working gain: on the signed build no gain lets the LIF pass all five core feeding and escape reflexes, while an adaptive LIF passes them from 0.46 to 0.53, against 0.43 to 0.49 on the reference build. One failure is shared by every build: driven through a connectome-constrained model of the eye, the escape pathway responds to a flash and not to a looming object. An independent analysis with a different signing rule reproduced the graph statistics and the loss of bitter suppression. Benchmarks of connectome models should state which graph they ran on. Code, data and the scripts that produce every figure and number: https://github.com/brandoncho369/flybench (paper/), archived as doi:10.5281/zenodo.23655868.