Fixed anatomy.
Plastic memory.

What comes from the connectome

Winding et al. (2023) reconstructed a first-instar Drosophila brain with 3,016 neurons and approximately 548,000 synapses. ButterflyBrain uses an induced subset of 32 neurons and 289 directed edges from Supplementary Data S1. Every edge retains its source ID, target ID, and observed synapse count.

One plastic connection class

Δwᵢⱼ = −η · rKCᵢ · rDANⱼ
wᵢⱼ ← clamp(wᵢⱼ + Δwᵢⱼ, 0.05, 1.00)

Only KC→MBON synaptic gains change. Dopamine gates depression during concurrent KC activity. FEED depresses the output assigned to avoidance; POKE depresses the output assigned to approach. The wiring never grows or rewires.

What is designed, not measured

Market features, KC_up/down/vol groupings, appetitive/aversive assignments, MBON valence labels, effective excitatory signs, LIF parameters, normalized output decoding, and body motion are engineering choices. Cell valences are not established by this EM dataset. The butterfly is a stylized visual identity. The learning circuit uses larval Drosophila wiring, not a butterfly connectome.

The engine steps LIF membrane dynamics 30 times at 10ms per tick. Spike counts are telemetry. Behavior reads normalized KC→MBON effective drive; it is not a validated spike-only motor model. DAN modulation follows matched MB compartments with observed DAN→KC contacts.

What ButterflyBrain never does

The larva learns valence. It does not predict price. It does not place orders. There is no screenshot vision, adult retina, trading wallet, launchpad automation, or order execution module.

Neuron map