FlyPool simulates a real fruit-fly brain: 138,639 neurons, mapped one by one under an electron microscope. When its sugar neurons fire, the treasury buys the token. When its escape neurons fire, tokens are burned. Nobody presses the button.
A decision that cannot be replayed is a decision that was invented.
Every decision ships with what you need to replay it on your own machine: same brain, same seed, same spikes, same action. If you get a different answer, the decision was false.
01 · In plain terms
01
The FlyWire connectome catalogues the neurons of an adult fly brain and their connections, reconstructed from thousands of electron-microscopy sections. We simulate 138,639 neurons and 15,091,983 connections, 40 % of them inhibitory.
02
Each neuron is integrated using the model of Shiu et al. (Nature, 2024), whose constants come from published electrophysiological measurements. Our implementation is checked against Brian2, the reference simulator: strict spike-for-spike equality, neuron by neuron.
03
It moves through an arena where sugar sources appear and run dry. It only eats if it reaches them. A visual threat looms over it and it flees. What it feels modulates what it perceives.
04
At regular intervals, a deterministic epoch measures one circuit and derives a bounded action from it. The seed is committed on-chain before the simulation runs: rerolling until the desired outcome appears is impossible.
02 · What the brain does to the token
The contract holds two piles: ETH (the treasury) and FLYPOOL (the reserve). Every 15 minutes during the first hour after launch, then every 30 minutes, the brain is measured. The circuit that fires decides what happens to one of those piles, and how much, within limits nobody can lift overnight.
Sugar → Buyback
ETH treasury → Uniswap pool → FLYPOOL out of the pool
A slice of the treasury's ETH buys FLYPOOL on the pool, at the pool's own price. The tokens leave the pool, the pool holds a little more ETH and a little less FLYPOOL, so the price edges up. There is no seller on the other side: only the pool.
Threat → Burn
FLYPOOL reserve → 0x…dEaD
A slice of the contract's FLYPOOL is sent to the dead address, a wallet nobody holds a key to. Those tokens can never move again: the supply that can circulate shrinks for good.
Grooming → Freeze
Decision recorded → no transfer
The decision is written on-chain with its proof, like any other, and deliberately moves nothing. It shows the brain was measured and chose to stay put.
Calm → Nothing
Below 5 % of a full response → no transaction
Background noise is not a decision. If the circuit fires at less than 5 % of a full response, no transaction is sent at all. Many hours look like this, and that is normal.
Nobody picks the moment, the size or the direction. The spikes do.
03 · Try a reading
This runs the published rule, the same one the engine applies: the firing rate is divided by 100 Hz to give an activity; under 5 % nothing happens; above that, the size grows with the activity until it hits its cap.
Buyback
An illustration of the rule, not a live quote. A real epoch publishes the measured rate and the exact size in its record, and the contract recomputes the amount from its own balances.
04 · Limits written in the contract
The engine proposes, the contract checks. Every limit below is enforced on-chain, and amounts are recomputed from the contract's own balances and the pool's reserves: a number sent from outside is never trusted.
2 days
The ETH treasury is spent evenly over the 2 days after the first decision, never faster.
15 → 30 min
Minimum wait between two epochs: 15 min in the first hour after launch, then 30 min.
1 % of the pool
Largest buyback, measured against the pool's ETH.
1 % a week
Circuit breaker on everything sent to the dead address.
0.1 %
Maximum share of the FLYPOOL reserve one epoch can burn; limits change only after a 2-day public delay.
Seed first
Committed on-chain before the brain runs, checked by the contract.
05 · Right now
The values below come from the live stream, not from a capture.
Hunger
-
State variable, documented constants
Feeding drive
-
Measured firing of motor neuron MN9
Neurons firing
-
In the last frame received
Last decision
-
No epoch received yet
Digest of the loaded connectome
-
06 · What holds, what does not
A project can claim it runs a brain while giving you nothing to check. Here is what we measured on the full connectome, and what we failed to establish. Both columns matter equally.
| Stimulation | MN9 firing rate | Share of brain active |
|---|---|---|
| 0 Hz | 0.0 Hz | 0.00 % |
| 20 Hz | 10.0 Hz | 0.21 % |
| 40 Hz | 59.5 Hz | 0.39 % |
| 60 Hz | 76.0 Hz | 0.39 % |
| 100 Hz (outside the domain) | 53.5 Hz | 6.39 % |
That last row is the reason every circuit declares an operating range that the code refuses to exceed. Past 60 Hz the response stops being monotonic and activity multiplies sixteenfold. A model is only trustworthy in the regime where its behaviour has been checked. We would rather publish that boundary than let anyone believe it does not exist.
| Claim | Status | What establishes it |
|---|---|---|
| The engine reproduces the published model | Verified | Strict spike equality with Brian2, four networks |
| A decision replays identically | Verified | Spike, protocol and decision digests all reproduced |
| A falsified record is rejected | Verified | Four tampering attempts tested, all detected |
| The escape circuit recruits the giant fiber | Verified | 0 → 78 → 105 → 137 → 178 Hz, under 1 % of the brain active |
| Actions execute on-chain | Testnet | On Robinhood Chain testnet, a real buyback: treasury ETH bought 47,833 test tokens on the Uniswap V2 pool; and a real burn: 25,000 test tokens sent to the dead address. Freeze verified on a local chain. Nothing on mainnet |
| The paper's quantitative calibration | Not reproduced | 80 % reported, 50 % measured, cause identified and recorded |
| The grooming command neuron responds | Disproved | DNg12 fires at no frequency; readout moved to DNg84 |
| The foraging circuit is usable | Disproved | 8 % of the brain active from 10 Hz; allocation stays unreachable |
| The contracts are audited | No | 36 tests including fuzzing and invariants, no external audit |
| Commercial use of the data is permitted | Unverified | Reference code licence: MIT. FlyWire data: not verified |
What FlyPool is not
It is not a living fly. The connectome is the map of one single individual, female, adult, taken post-mortem. The model covers the central brain and excludes the ventral nerve cord. It has neither plasticity nor learning: the fly does not improve.
These are not emotions. They are persistent internal states documented in the literature, displayed under their scientific names, with their citations. "Defensive arousal", not "fear".
The mapping from neural activity to token action is a convention we chose. Biology determines the input, not the financial decision. Not conflating the two is the only thing we ask.
07 · Nothing here asks to be believed
The connectome snapshot is frozen and identified by its digest, checked on every load. The engine code is identified by its commit. Each epoch's seed is committed on-chain before the computation and revealed afterwards: picking the seed that yields the desired action is impossible, and the contract verifies this itself.
A single command replays any published decision on your own machine, from its archived record. It recomputes the spike trains and compares the digests. If they differ, the decision was falsified, and you needed no one's word for it.
We are not asking you to believe us. We are giving you what you need to redo the computation.
08 · Every claim, and where it comes from
Nothing here is ours to assert. Each line below is a published paper, with its DOI: the connectome we run, the model we reproduce, and the measurements behind every internal state we display. What is ours, the arena, the state constants, the mapping from neural activity to an on-chain action, is named as such, in the method.
| Authors | Paper | What it establishes |
|---|---|---|
| Dorkenwald et al.Connectome | Neuronal wiring diagram of an adult brain Nature 634:124-138 (2024) doi:10.1038/s41586-024-07558-y | The map itself: 139,255 proofread neurons, ~54.5 M synapses. |
| Schlegel et al.Connectome | Whole-brain annotation and multi-connectome cell typing of Drosophila Nature 634:139-152 (2024) doi:10.1038/s41586-024-07686-5 | Cell types. This is what lets a circuit be addressed by type rather than by an opaque identifier. |
| Eckstein et al.Connectome | Neurotransmitter classification from electron microscopy images Cell 187:2574-2594 (2024) doi:10.1016/j.cell.2024.03.016 | The sign of each connection, excitatory or inhibitory. Without it no simulation is possible. |
| Shiu et al.Model | A Drosophila computational brain model reveals sensorimotor processing Nature 634:210-219 (2024) doi:10.1038/s41586-024-07763-9 | The leaky integrate-and-fire model this engine reproduces, with no parameter tuned per circuit. |
| Gibson et al.States | Behavioral responses to a repetitive visual threat stimulus express a persistent state of defensive arousal Current Biology 25:1401-1415 (2015) doi:10.1016/j.cub.2015.03.058 | Defensive arousal persists for minutes and accumulates with repetitions. |
| Inagaki et al.States | Visualizing neuromodulation in vivo Cell 148:583-595 (2012) doi:10.1016/j.cell.2011.12.022 | Food deprivation raises gustatory sensitivity to sugar. |
| Root et al.States | Presynaptic facilitation by neuropeptide signaling mediates odor-driven food search Cell 145:133-144 (2011) doi:10.1016/j.cell.2011.02.008 | Hunger modulates odour-guided food seeking; the complement we display as satiety. |
| von Reyn et al.Circuits | A spike-timing mechanism for action selection Nature Neuroscience 17:962-970 (2014) doi:10.1038/nn.3741 | The giant fiber pathway and the escape decision it carries. |
| Card & DickinsonCircuits | Visually mediated motor planning in the escape response of Drosophila Current Biology 18:1300-1307 (2008) doi:10.1016/j.cub.2008.07.094 | Planned versus unplanned take-off, the behaviour behind the burn mapping. |
| Klapoetke et al.Circuits | Ultra-selective looming detection from radial motion opponency Nature 551:237-241 (2017) doi:10.1038/nature24626 | LPLC2 as the looming detector we stimulate for the escape circuit. |
| Hampel et al.Circuits | Distinct subpopulations of mechanosensory chordotonal organ neurons elicit grooming eLife 4:e08758 (2015) doi:10.7554/eLife.08758 | Antennal mechanoreception drives grooming, the circuit walking stimulates here. |
| Guo et al.Circuits | A neural command circuit for grooming movement control Current Biology 32:339-353 (2022) doi:10.1016/j.cub.2021.12.055 | DNg12 as the grooming command neuron. In our runs it stays silent, a negative result we publish. |