Connectome controls

The problem

I built a webcam demo where a fly dodges your hand using published escape wiring from a fruit fly connectome. It worked. I could not tell whether the wiring was doing the work, or whether anything of roughly that shape would produce a dodge.

The system

Three controls. A degree-preserving shuffle that keeps every neuron’s exact in and out degree and permutes only who connects to whom. A targeted ablation of the loom-sensitive cell types. And a trained readout on the frozen network, benchmarked against raw downsampled pixels with matched feature counts and held-out backgrounds.

What happened

The wiring carries real structure that shuffling destroys, significant on all three tasks. And it still never clearly beats raw downsampled pixels. Both are true at once.

Not claiming

One visual system’s published wiring plus its descending targets, simulated. Not an organism, not validated against physiological recordings. The looming signal is computed in software and injected, so what is tested is the visual-to-motor step rather than the detection.