A real fruit-fly connectome works a five-minute shift at a desk. What you watch is a recording, replayed step for step: the official MaleCNS v1.0 connectome — 166,700 neurons and 25,582,938 directed edges — simulated as an alpha-synapse LIF conductance model at 0.1 ms on a native sparse C++ backend. The shift ran 300 seconds of brain time in one continuous take.

The task

The fly watches a CRT. Five fruit appear as gaze targets; each needs a 200 ms fixation to refine. The same 36 boards run twice, in paired halves: part 1 calm, then part 2 with the swarm — 72 episodes, all solved. The seed list restarts at the chapter boundary, so every board is a control/distraction pair.

A perceiving servo

Steering is a scripted visual servo, not an omniscient expert. While a swarm fly does not cover the fruit, it reduces exactly to the old oracle — control-half state and action trajectories are byte-identical to that oracle on the same seeds. When a fly disk occludes the target in the eye's equirect angular space (depth gated), the servo freezes its belief and drifts. If the overlap holds for a run of steps, it fly-locks: gaze is pulled onto the passing face, then recovers. Geometry triggers the error; the response constants are tuned.

What happened

Paired solve time rose a median of +7 steps (IQR 0 / +56). Boards with more than 30 occluded steps were +45 steps slower on average. Twenty-five of the 36 pairs ran longer in the swarm half. The servo fly-locked 43 times and reset dwell 5 times under distraction, versus none in the calm half.

Effect size is a function of those controller constants — inspect rate, overlap thresholds, lock window — tuned on a disjoint pilot seed set. Only the swarm's geometry and timing are emergent. Part 2 boards are also solved by a ~150-second-older brain; treat the paired numbers as descriptive of this recording.

Interesting observations — yes, there are some

The connectome's descending neurons track the motor command underway rather than the visual error remaining. A supervised linear controller fit them beautifully on expert paths (R²≈0.96) and failed closed-loop everywhere else (9/9 evals, 0/5 targets) — recorded as decision D-042 in the project. That negative result is unchanged.

Honest framing: steering is still a script — now a perceiving script whose errors come from real swarm geometry. The brain's response is fully computed, not faked. The flies on screen are the same perch/cruise/inspect swarm the eyes received; the v1 display that quietly disagreed with the recording is gone.

Credits

Danny Halarewich — halarewich.com, LinkedIn. MaleCNS/FlyWire connectome. Built with Three.js.