FLYPIT documentation

FLYPIT is a futures desk whose only decision maker is a simulated fruit fly brain. This document explains exactly what is simulated, exactly how a price series becomes an order, and exactly what we cannot promise.

What FLYPIT is

A token on Robinhood Chain taxes its own trading. The tax funds a wallet. That wallet trades a single perpetual futures market. The direction and size of every trade come from a simulated fruit fly connectome that has been shown the market as visual motion. No human is in the loop after launch, and no human can withdraw from the desk.

There are three moving parts:

1. The token. Standard tax-on-transfer token. Split is fixed at deploy.
2. The signal service. Ingests price, renders the visual field, runs the circuit, emits a steering scalar.
3. The relay. Takes the steering scalar, applies risk rails, sends orders, publishes the log.

Status and honesty

We would rather ship a smaller honest claim than a large unfalsifiable one, so here is the boundary. We run the visual and descending motor pathways from the released connectome. We do not run a whole animal, we do not claim consciousness, and we do not claim the fly understands markets. It does not. It responds to motion, and we have chosen to treat that response as a trading signal, which is a bet we are making in public with a public balance.

The connectome

In September 2026 Google Research and HHMI Janelia released a whole nervous system reconstruction of Drosophila melanogaster: roughly 166,000 neurons and about 125 million synaptic connections spanning the central brain, both optic lobes and the ventral nerve cord. It is the first complete wiring diagram of a nervous system that can both see and walk.

We use three regions of it:

Everything else in the dataset is loaded but not driven. Olfactory, gustatory and mechanosensory inputs are held at baseline because we have nothing honest to feed them. A market has no smell.

Encoding price as flow

This is the only place where we make a creative choice, so it is the part worth scrutinising.

A fly's visual system is built to answer one question extremely well: is the world sliding past me, which way, and how fast. That is optic flow. We hand the market to the fly as optic flow and nothing else.

The transform

// 250 ms tick, 480 sample ring buffer = 2 minutes of mark price
v  = ema(dPrice/dt, halflife=6s)        // price velocity
s  = stdev(returns, window=120)          // realised vol

flowDegPerSec = clamp(v / (s * k), -180, +180)   // k = 2.5
contrast      = clamp(0.30 + s * 14, 0.30, 0.95)

frame = squareGrating(
  width      = 720, height = 360,            // 1 px ~ 0.5 deg azimuth
  spatialFreq = 1/22 cycles/deg,
  drift       = flowDegPerSec,
  contrast    = contrast
)

Three things are deliberate here:

• Velocity is normalised by volatility. A two dollar move in a dead tape and a two dollar move in a violent one are not the same event, and the fly should not see them as the same speed.
• Contrast carries volatility. In a real fly, higher contrast produces a stronger and faster HS response. So a volatile tape produces a more decisive fly, which is the behaviour we want and it falls out of the biology rather than being bolted on.
• Spatial frequency is fixed near the peak of the measured HS tuning curve. We are not tuning this parameter to fit returns. It is set from the literature and left alone.

Running the circuit

The frame is projected onto the photoreceptor array and propagated through the loaded connectivity. Neurons are leaky integrate and fire with synaptic weights taken from reconstructed synapse counts and sign assigned by neurotransmitter prediction from the dataset.

This is a coarse model. Real neurons are not uniform leaky integrators and the release does not hand us biophysics, only wiring. We are not going to pretend that gap does not exist. What the wiring does give us is the thing that matters for direction selectivity: the specific asymmetric delay lines that make a fly's motion detector work. That structure is in the data and it is what produces the left right asymmetry we read.

Decoding steering

We read the descending neurons associated with turning, sum them by side, and take the difference.

steer(t) = sum(rateL) - sum(rateR)          // per 250 ms tick
bias     = mean(steer, window = 120 ticks)  // 30 seconds
conv     = clamp(|bias| / B, 0, 1)          // B calibrated on grating sweeps

if conv < 0.35            -> FLAT
else if bias > 0          -> LONG
else                      -> SHORT

notional = equity * 0.35 * conv
leverage = min(3.0, 1.0 + 2.0 * conv)

B is calibrated once, by sweeping the simulated fly with known grating speeds and recording the steering response. It is a property of the circuit, not of any price series, and it is not refit against returns. That distinction is the whole difference between this and a curve fitted strategy wearing a fly costume.

Why a 30 second window

Raw steering is extremely twitchy, because a real fly corrects its heading many times a second. Traded raw it would pay the spread to death. Thirty seconds is long enough to suppress correction reflexes and short enough to keep the response that gave the fly its direction in the first place. It is a judgement call and it is the parameter we are least certain about.

Determinism and replay

There is no stochastic term anywhere in the pipeline. Same price window in, same steering out, on any machine, forever.

Every tick we publish the input hash and the output:

{
  "t": 1789200000000,
  "inputHash": "sha256:9f2c…",   // hash of the 480 sample price window
  "flow": -37.4,
  "steer": -0.181,
  "bias": -0.094,
  "conv": 0.508,
  "decision": "SHORT",
  "simRev": "flypit-sim@1.0.3"
}

Anyone can pull the order log, re-run the published simulator at the pinned revision against the published price window, and get the same decision. If they do not, that is a bug report we want.

Risk rails

The fly has no concept of ruin. The rails are the only adult in the building.

The relay

The relay is the only component that holds an execution key. It is deliberately small and deliberately stupid.

• It accepts a decision object signed by the signal service.
• It verifies the signature, the simRev, and that the input hash is fresh.
• It applies every rail above, in order, and records what each rail changed.
• It sends one order. It cannot send a batch, and it cannot open a second market.
• It has no withdrawal function. The key it holds can trade and can call the buyback contract. That is the entire capability surface.

Public order log

Every order, every clamp and every rail trip is appended to a public log, published at /api/log and mirrored to chain in daily Merkle roots so the history cannot be quietly edited.

GET /api/log?from=1789200000000&limit=100

{
  "orders": [
    { "t": …, "side": "SHORT", "notional": 412.88, "lev": 2.0,
      "conv": 0.51, "clamped": false, "txHash": "0x…" }
  ],
  "merkleRoot": "0x…",
  "anchorTx": "0x…"
}

Fee routing

4 percent on buys and sells, split at the contract level, no privileged setter after renounce.

Buyback and burn

When a position closes in profit, the realised gain splits in half.

• 50 percent is used to buy $FLYPIT on the open market and send it to the burn address.
• 50 percent stays in the desk and compounds.

Losses do nothing. There is no burn on a losing trade, no burn on a schedule, and no burn from a reserve we quietly held back. If you see a burn, a trade paid for it. That is the only mechanism.

Contracts

Addresses appear here and on the front page at the same moment they go live. There is no private phase where they exist and are not listed.

HTTP API

GET /api/desk

{
  "configured": false,
  "reason": "DESK_TREASURY unset"
}

Known limitations

Read this section before you read anything else on the site.

1. Wiring is not biophysics. The release gives us connectivity, not per neuron dynamics. Our uniform leaky integrator model is a strong simplification and a different plausible choice of parameters would produce a different fly.
2. The encoding is our invention. The circuit is real. The decision to show it a price series as a drifting grating is ours, and it is unfalsifiable as a piece of biology. It is defensible, it is not proven.
3. No edge is claimed. We have not published a backtest and we are not going to, because a backtest of a hand designed encoding over one asset proves nothing and would be used to imply a track record we do not have.
4. One market, one position. No hedging, no basket, no market selection. If the traded market goes quiet, the desk goes quiet.
5. The relay key is hot. Compromise means an attacker can lose the desk money, though not take it.
6. Fee income depends on volume. If nobody trades the token, the desk never grows past its starting balance, and a buyback funded by a small desk is a small buyback.

Glossary