We ran a fly's escape reflex on a desktop

What was built

The release contains 151.8 million connections across 87.5 million bodies — but most of those bodies are unproofread fragments. Restricting to traced, annotated neurons gives the network we actually ran:

The model is deliberately crude: leaky integrate-and-fire. Each neuron sums its signed input, fires past a threshold, resets. Weights are synapse counts; sign comes from the presynaptic transmitter. No conductances, no dendrites, no neuromodulation, no plasticity, no body. If a signal still arrives where it should under these assumptions, the structure carried it.

Test one: the panic button

The Giant Fiber — DNp01 — is the most studied neuron in any brain. One per side, unusually thick axon, and one message: go now. It is the fly's equivalent of the Mauthner cell that produces a fish's escape flick.

We stimulated the two Giant Fibers for three steps and let the network run.

Two neurons reach 19% of the brain and most of the motor population, first motor spike at step three. Scramble who connects to whom — keeping every neuron's exact weights and degree — and the same stimulation dies immediately. Stimulate two random neurons anywhere in the brain and nothing happens.

Test two: the prediction we got wrong

In a living fly, two visual populations — LC4 and LPLC2 — detect looming and converge on the Giant Fiber. Other visual populations target different descending neurons. So, staked before running:

PREDICTION: LC4 + LPLC2  -> the Giant Fibre fires
            LC6 + LC22  -> it does not

It failed. Every population fired everything. LC6 and LC22 drove the Giant Fiber just as reliably as the looming detectors, and each stimulation ignited about 21,000 neurons.

The cause was not biology but our own dynamics: with a hundred-odd simultaneous inputs, excitation snowballed and the whole network went off regardless of the channel. Real brains hold excitation and inhibition close to cancelling, and ours had no such constraint.

The fix, and the confirmation

One change, no new data and no invented biology: normalise each neuron's total excitatory input, so every cell has a comparable operating point instead of inheriting whatever its raw synapse count happens to be. Inhibition is scaled by the same per-neuron factor, so each cell's excitation-to-inhibition ratio is preserved.

The prediction then holds — and five controls say it is specific:

Larger populations do not do it. Size-matched populations do not do it. Random draws do not do it. Only the two populations that converge on the Giant Fiber in a living fly.

What the wiring could not do

In the balanced regime the escape command reaches the Giant Fiber and stops there — zero motor neurons, at every gain that preserved selectivity. Turn the gain up until motor neurons are reached and selectivity collapses: every channel fires everything again.

Selectivity or propagation. Not both.

That is a measurement rather than a disappointment, and it is the number we would most want a reader to take away. A connectome plus the simplest possible dynamics gives you a command pathway and a correct sensory trigger. It does not give you a working nervous system, and the gap is not compute — it is the physiology the wiring diagram does not contain: conductances, timing, adaptation, and the neuromodulatory state that reconfigures circuits wholesale.

What this does not settle

• It is not news to fly neuroscientists. That LC4 and LPLC2 drive the Giant Fiber is established. The point here is that it falls out of the public wiring plus eighty lines of arithmetic — and that the naive version fails without balance.
• "Motor neurons reached" is not "the fly jumped." No muscles, no body, no world.
• Only 118 motor neurons are annotated in this release, and the classic jump-muscle targets are not among them by name.
• Synaptic sign is inferred from the sender's transmitter. The true sign depends on the receiver's receptors, which this dataset does not contain. Some fraction of our 17 million edges therefore carry a sign that is a guess about the wrong end of the synapse.
• Steps are not milliseconds. No timing claim is made.

Reproduce it

BASE=https://storage.googleapis.com/flyem-male-cns/v1.0/connectome-data/flat-connectome
curl -O $BASE/body-annotations-male-cns-v1.0-minconf-0.5.feather      #  14 MB
curl -O $BASE/body-neurotransmitters-male-cns-v1.0.feather            #  42 MB
curl -O $BASE/connectome-weights-male-cns-v1.0-minconf-0.5.feather    # 1.0 GB

python build_matrix.py     # 14 seconds
python escape.py           # the Giant Fibre test, with its shuffle control

Total cost: one download and a few minutes of CPU. No model calls, no cloud, no funding.