Why the model is grounded.

Connectomics reconstructs complete wiring diagrams: every neuron, every synapse, the direction and strength of each connection. For the adult fruit fly this is finished work. The FlyWire consortium published a whole-brain connectome of 139,255 neurons and roughly 54.5 million synapses in 2024, following the hemibrain reconstruction of the central brain in 2020.

Completeness changes what a model can be. A network built on a connectome does not assume its connectivity, which narrows the modelling choices to the neuron model and a small number of free parameters.

Biological neurons communicate in discrete, timed events. A spiking model keeps that structure: state evolves continuously, but information crosses synapses only at spike times. The temporal organisation of the computation stays explicit instead of being averaged away.

NeuroSim uses leaky integrate-and-fire neurons, the model used by the published whole-brain Drosophila work whose parameters it adopts. It captures integration, leak, threshold, reset and refractoriness. It does not model dendritic computation, active conductances, short-term plasticity or neuromodulation.

A fruit fly navigates turbulent odour plumes, learns associations in a handful of trials and tracks a heading through featureless terrain, on a few hundred microwatts. Whatever intelligence means at that scale, it is implemented by about 139,000 neurons whose wiring is now known.

The question is narrow and answerable in principle: how much of that capability is determined by circuit structure. Where structure alone predicts function, the connectome is close to a specification. Where it does not, the gap identifies what else the model needs.

A circuit studied in isolation is fed inputs an experimenter chose. A circuit in a loop with a body receives inputs its own past actions produced. That difference changes what the dynamics mean.

NeuroSim closes the loop at the smallest scale that is still honest. An agent's two visual half-fields are computed by ray-casting an arena, injected into the two halves of the optic-lobe circuit, and the resulting readout rates set the two wheel speeds. The body is a kinematic model and makes no claim to reproduce fly locomotion.

Biological circuits are sparse, event-driven and asynchronous. Neuromorphic engineering asks what happens when hardware adopts those properties, and connectome-derived circuits are natural benchmarks for it.

The accelerated backend is a small exercise in the same direction: a spiking network as compute passes per timestep, with delayed delivery through a ring buffer. It is a browser kernel rather than neuromorphic hardware, but the structure of the problem is the one GPU spiking simulators solve.