I built this for my kids, then kept going because the honest version turned out to be the interesting one.

The data is MaleCNS v1.0 (Berg et al. 2026, CC BY 4.0), the male fly central nervous system. 166,691 neurons, and 140,024 of them have measured soma coordinates. Those are the dots you see, at their real positions.

The activity on top is a leaky integrate-and-fire simulation I ran offline, with the parameters from Shiu et al. 2024. Chemical synapses only, because that is what a connectome records. No plasticity either: the weight table is read only once it is loaded.

What that costs, all of it stated on screen rather than in a footnote:

The visual front end does not carry the signal. Driving the optic columns gets LC4 to 1.23 Hz and LPLC2 to zero, so the looming input is injected. It is drawn in magenta and labelled as ours.

The model never stops. In 20 runs to 600 ms, not one settled on its own. Playback ends at 250 ms and the wind-down after that is drawn by me.

PSI, the relay the textbook puts between the giant fiber and the wing depressor, never fires here. I ran an ablation: zero out its inhibitory inputs and it fires in all three seeds, so inhibition is what holds it down. But the wing readout does not change, and in the fly that GF to PSI synapse is mixed electrical and chemical, with the electrical half dominant in this circuit. A connectome has no electrical synapses in it, so this model cannot test that route. Its silence is not evidence against the textbook.

Every number on a card carries a mark: computed by a script in the repo, read first-hand from a paper, or secondhand. Ten deterministic gates check those marks, check that retracted wording never comes back, and check that no Chinese is left on the English pages. When I asked a model that had not helped build any of it to attack those gates, it got 25 obviously wrong strawmen past every one of them. That is in the repo as well.

The engine and the weighted connectivity file live in a private research repo, so you cannot re-run the simulation yourself. The raw per-run outputs are public, so every median and range on the site is recomputable from what is there.

Site: https://fly-brain-explorer.vercel.app Lab (swap the stimulus set, 9 conditions run 20 times each): https://fly-brain-explorer.vercel.app/?lab=1 Code and audits: https://github.com/klmtseng/fly-explorer

The English has not been reviewed by a native speaker. Corrections to the science are the most welcome kind of comment.