FLY STREET: DATA, CODE AND MODEL BOUNDARIES

MaleCNS v1.0 by the FlyEM Project Team (HHMI Janelia Research Campus), University of Cambridge, MRC Laboratory of Molecular Biology, and Google Research. Data licensed CC BY 4.0: https://creativecommons.org/licenses/by/4.0/
Publication: Berg et al., Sexual dimorphism in the complete connectome of the Drosophila male central nervous system, Cell (2026). https://doi.org/10.1016/j.cell.2026.08.015
Dataset: https://male-cns.janelia.org/

SIMULATED CIRCUIT
Unmodified locomotor_circuit.json from DenisSergeevitch/desktop-fly commit 32b00011e83c3dc85fa3ea0b3934155b04f1635d.
1,045 neurons (16 descending, 622 premotor, 220 motor, 153 sensory, 34 ascending), 17,224 directed connections.
Source extraction rules and measured-versus-modeled boundary: data/LOCOMOTOR_PROVENANCE.md
Data license: data/DATA_LICENSE.md (MaleCNS supplement applies). No FlyWire-derived data is bundled in Fly Street.
The server and browser replay run the original LocomotorSim leaky integrate-and-fire engine at 1 ms neural steps. Scan ticks process 20 neural steps. Its physiological parameters, synaptic normalization and neurotransmitter sign choices are modeling assumptions. The browser replays persisted scan inputs with slowed visual timing. An overloaded browser slows replay rather than changing the shared ledger.
Live market inputs stimulate DNp09 (positive momentum) and MDN (negative momentum plus volatility). The same 1,045-cell graph is scanned for 500 ms per asset, with 100 ms startup discarded; circuit state resets between assets. The current universe is SOL, cbBTC, TSLAx, NVDAx, SPYx, GOOGLx, AAPLx, MSFTx, AMZNx and METAx. Each scan models five seconds of neural time across ten assets. The mean firing rate of left hind-leg motor neurons is the buy readout; left front-leg motor neurons form the sell readout. Score = buy Hz - 3 * sell Hz. Buy at score >= 5, sell at <= -3, otherwise hold. The strongest eligible sell for an existing position has priority; otherwise the strongest qualifying buy wins. Input encoding, output assignments, factor 3, thresholds and cost assumptions are hand-designed. No reward learning or predictive advantage is claimed. Leg animations are cosmetic; body mechanics do not execute trades. The sensory body-feedback loop is not enabled.

WHOLE-CNS CONTEXT
context.json is every eighteenth finite soma coordinate from blendi-remade/fly-brain-minecraft commit 6cfa30175003ef25da68a237d5eda958f8047b82, file src/main/resources/connectome/malecns-v1.0.flyb.gz.
Original file SHA-256: e33df182bed7a6f3ea279daf4790a82b05706d3d41e819a6a80c0473e8c559f3 (verified before extraction).
7,877 sampled coordinates, rounded to whole voxels, projected using source x and z. Background samples provide anatomy context only; they are not simulated or assigned fake firing activity. Active neurons are plotted from the locomotor circuit's own measured soma coordinates.
Full upstream provenance: data/MINECRAFT_PROVENANCE.md. The mod includes 176,422 database Neuron nodes; the publication describes 166,691 proofread neurons. Neither total describes the number simulated by this game.

CODE
desktop-fly simulation and procedural Three.js fly body: MIT, copyright 2026 Denis Shiryaev. License: vendor/DESKTOP-FLY-LICENSE.txt. The flymodel import path is changed for browser hosting. The neural engine adds per-neuron spike counters for visualization; its model equations are retained.
fly-brain-minecraft: MIT, copyright 2026 Blendi Remade / fal.ai. License: vendor/MINECRAFT-LICENSE.txt. Used as the source of the context data and provenance, not as the game's executable engine.
Three.js 0.169.0: MIT, license vendor/THREE-LICENSE.txt.

THE MILLION DOLLAR EXPERIMENT

Ten selected Solana-traded assets, verified against Jupiter Tokens V2 on 2026-09-10. This is a selected universe, not a ranking. xStocks are tokenized exposures, not direct exchange-listed shares; cbBTC is Coinbase wrapped Bitcoin. References: https://xstocks.com/products and https://www.coinbase.com/cbbtc . The public manifest records exact mints in data/experiment-assets.json.

Quotes: https://api.jup.ag/price/v3 . The shared service samples at most every 30 seconds while followers request the feed. USD is the accounting unit. No scheduler runs between visits. Missing observations are never backfilled as trades. Six advancing source-block observations are required per asset. A block that has not advanced in 120 seconds pauses trading; observation gaps reset history. First observed block age is not independently established, and source prices are not executable swap quotes. Fetch failures preserve the last ledger and display a delayed state.

The portfolio begins with $1,000,000 hypothetical cash. Every order is a modeled fill, not a blockchain transaction. Version 2 buys spend at most 2.5% of current equity, with a 12.5% entry cap per asset and 60% total exposure ceiling for new buys. Eligible sells exit the position. Ordinary trades require fifteen-minute asset cooldowns and buys are at least two minutes apart. Estimated fees are 0.10% and slippage 0.05% per side. Liquidity, market impact, taxes and corporate actions are not modeled. No leverage or shorting is used.

The benchmark starts at the first complete observed price set and buys equal gross dollars of all ten tokens, applying the same entry-cost assumptions. Portfolio P&L includes cash, realized gains, unrealized gains and modeled fees. Last-price valuation persists during data gaps. The ledger is shared and durable in D1, with compare-and-swap revision checks preventing duplicate allocations across viewers. It keeps lifetime totals, the latest 200 trades, and a bounded multi-resolution inception-to-date chart.

Version 1 used short price histories with momentum scaled by 0.005. Version 2 uses timestamped five- and twenty-minute windows. Positive movement must exceed a hurdle of at least 1.5 times modeled round-trip friction and 1.5 times observed return standard deviation times the square root of observations. Directional efficiency must be at least 0.35. Two same-direction neural responses on distinct source blocks are required. A 2.5% observed position drawdown stimulates protective avoidance that can bypass normal sell cooldown/confirmation only when the neural output produces a sell. An 8% portfolio drawdown suspends new entries. Stops are not guaranteed fills during quote or observation gaps. The measured connectivity is not trained on financial returns. Fixed readout rules translate motor rates into buy, sell or hold; a neural score alone is not evidence of predictive value. Identical recorded inputs are replayed through the same deterministic model in the browser. Sparks originate from actual model threshold crossings on measured connections, with stylized travel time. The fly body visits the corresponding assets and dwells on its recorded choice. This is a locomotor-response experiment, not a biological taste model or the complete fly brain.

Earlier vault and wallet development source is retained separately but excluded from the public experiment build. No real funds are requested, held or traded.

LAUNCH RUN
The public launch has its own run identity, $1,000,000 initial accounting balance, zero initial positions and new performance timestamp. The prior development run remains archived separately. The launch uses the revised strategy but does not run or display the previous-rule comparison. No real funds are held or traded.

Asset brand icons: Simple Icons 11.15.0, CC0. https://simpleicons.org/
S&P 500 uses a typographic index marker. Brand marks identify the assets, without implying endorsement.