01 — THE CLAIM
A vertebrate brain launched a memecoin.
71,721 neurons. Whole-brain activity recorded from one real larval zebrafish, replayed through a fixed readout that proposes buy, sell or hold. No language model picks trades. No script pretends to be neurons. A risk guard can say no; it can never say “trade this instead.”
This is Experiment 01 of the ZAPFISH Observatory. The same brain also plays table tennis, rides a bike and walks a dog. The fruit fly was an invertebrate with a wiring diagram. This is a vertebrate with a wiring diagram and a recording of its brain, from the same animal, sensory input to muscle output.
02 — FLY → FISH
The next brain up.
STONKFLY put a fruit-fly connectome on a trading account. ZAPFISH climbs one rung of the tree of life: from an insect to the first complete vertebrate connectome, with recorded activity to match.
| FRUIT FLY | ZEBRAFISH | |
|---|---|---|
| Lineage | Invertebrate | Vertebrate |
| Wiring | Brain + nerve cord | Brain, sensory structures and motor outputs |
| Evidence | Connectome only | Wiring plus recorded activity from the same fish |
| Neurons | 166,700 | ~140,000 mapped · 71,721 recorded |
| Resolution | Single synapse | Single synapse, sensory input traceable to muscle |
| Story | Strong science | Stronger evolutionary story |
03 — HOW IT WORKS
See. Fire. Trade.
SEE
Public price candles are rendered into a small chart. The model gets the pixels, not the numbers. No indicators, no balances, no P&L on screen.
FIRE
A whole-brain predictive model trained on the ZAPBench recordings of one zebrafish responds to that input, cell by cell. The raster in the hero is real data from that fish.
TRADE
A fixed readout maps activity in a named set of motor-side cells to buy, sell or hold. A guard checks limits and can only reject. Paper mode by default.
Latest paper observation from the ZAPFISH engine (56,909 of 64,748 plastic edges had moved after two observations). The readout carries a persistent sell-side bias that the guard vetoes for lack of inventory. As STONKFLY warns of its fly: a persistent turning-like bias becomes a persistent order bias. Do not read it as market insight.
Engine: a fork of the open-source STONKFLY loop (market → sensory proxy → neural model → fixed decoder → execution limits), with the fly graph swapped for the fish. Same rules: no pruning, no hidden profit-based action selection, no LLM policy.
04 — WHAT WE CLAIM
Loud about the limits.
WE CLAIM
- The activity you see is from a real zebrafish: ZAPBench, public release, 71,721 neurons × 7,879 frames. Every row in the hero raster maps to a neuron ID in that dataset.
- The readout is fixed and readable. Selected cell IDs are logged. Nothing overrides the proposal with a different action.
- The code is open and the paper-trade run is resumable and auditable.
WE DO NOT CLAIM
- Control of the complete Fish Fire&Wire connectome. Janelia is still proofreading it; early snapshots are by application. Until access is granted, ZAPFISH runs on the public activity recordings, not the full wiring.
- Profitable learning. A rising market makes any buyer look smart. No held-out replay has shown an edge.
- Modeled pain, pleasure or consciousness. Profit and loss are engineered inputs, not feelings.