The model
A connectome is a wiring diagram. It records which neurons contact which others and how often. It does not record what any of them do. Everything between the diagram and a fly that walks is an assumption, and this page lists ours.
The graph
The source is the MaleCNS v1.0 reconstruction of an adult male Drosophila melanogaster — the complete brain and ventral nerve cord, 166,000 neurons and roughly 125 million synaptic connections, released under CC-BY. Annotations and connectivity are queried through the neuPrint API against dataset male-cns:v1.0. Neuron identifiers are the released identifiers and are never renumbered.
We do not simulate all of it. The retained subgraph is the central complex — the ellipsoid body, protocerebral bridge, fan-shaped body and noduli — together with the lateral accessory lobes that carry its output, the visual pathway that feeds it, the descending neurons that drive locomotion, and the giant fibre DNp01. A connection is kept only at or above a synapse-count threshold, which discards the long tail of one- and two-synapse contacts where reconstruction error concentrates.
The retained counts are a property of the import, so they are reported by the import rather than asserted here. Until it has run against the released graph they stay empty.
| Neurons retained | not yet measured |
|---|---|
| Connections retained | not yet measured |
| Neuropils included | not yet measured |
| EPG wedges | 16 |
Dynamics
Each retained neuron is intended to be a single-compartment leaky integrate-and-fire unit: one membrane potential, no dendritic tree, no axonal delay beyond a fixed synaptic time constant, no channel kinetics. It integrates an exponentially decaying synaptic current, crosses threshold, emits a spike, resets, and is silent for a refractory period.
That kernel is not running. What is running is described on what is running, and the parameters below belong to the kernel, so they are empty rather than filled with numbers from somewhere else.
| Parameter | Value |
|---|---|
| Integration step | not yet measured |
| Membrane time constant | not yet measured |
| Synaptic time constant | not yet measured |
| Resting potential | not yet measured |
| Spike threshold | not yet measured |
| Reset potential | not yet measured |
| Refractory period | not yet measured |
| Weight per synapse | not yet measured |
| Synapse retention threshold | not yet measured |
The eye
The arena reaches the model only as light. Terrain is rendered from the fly’s own position and heading into a photoreceptor response, which feeds the visual pathway, which feeds the ring neurons of the ellipsoid body. Ring neurons inhibit, so a bright direction suppresses the part of the heading ring facing it — the same pathway that lets a real fly hold a course against a landmark. Here the landmarks are the mesas.
| Ommatidia | not yet measured |
|---|---|
| Azimuth bins | not yet measured |
| Elevation bins | not yet measured |
| Acceptance angle | not yet measured |
Sign
A wiring diagram carries no signs. To run one you have to decide, for every neuron, whether its output excites or inhibits, and the connectome does not say. We use the transmitter predictions released with the dataset: acetylcholine is treated as excitatory, GABA as inhibitory, and glutamate as inhibitory, on the grounds that the glutamate-gated chloride channel is widespread in the fly central brain. Every synapse a neuron makes carries that neuron’s sign.
This is a coarse proxy, and it is wrong in ways we can already name. It assumes one transmitter per neuron, which co-transmission violates. It assigns sign by transmitter rather than by receptor, so the same released molecule cannot be excitatory at one target and inhibitory at another, which in a real fly it is. It contains no neuropeptides and no neuromodulation, so there is no internal state: no satiety, no arousal, nothing that would make the same input produce a different response an hour later. And a prediction that is wrong for a given neuron flips that neuron’s entire output, not part of it.
What this is not
The connectome supplies anatomy. It does not supply a calibrated living fly, and assembling a plausible simulation on top of it does not produce one. The distance between this model and the animal is made of receptor identity, channel distribution, dendritic computation, neuromodulatory state, the body, and the world. None of those are in the graph, and none of them are in here.
Nor is any of it a strategy. Nothing in this model was selected, tuned or trained against price. It was assembled from a published anatomy and a list of stated assumptions and then pointed at a market to see what it does. Whether what it does is worth anything is the subject of validation, and the answer there is currently that nothing has been shown.