Paper-trading research desktop that drives the full MaleCNS fruit-fly connectome with crypto market data.
Warning
Experimental research software. Coinfly is shared as-is for research and education. There is no guarantee that it works, is stable or will be updated, and no claim that it trades profitably. Paper execution is the default and live trading is closed. This is not financial advice. Use it at your own risk.
Coinfly feeds crypto order-book data into the complete retained MaleCNS v1.0 male fruit-fly connectome (166,700 neurons, 25,582,938 edges, no pruning) and lets identified descending neurons propose paper trades. A separate worker process owns the autonomous loop, exact accounting, risk limits and isolated paper execution; the desktop app only sends commands and shows status.
- The whole graph, no shortcuts. A native C++ spiking kernel runs every retained neuron. A HOLD is never replaced by a scripted trade, and no LLM policy or hidden profit rule selects actions.
- Exact, durable accounting. Decimal quantities, order intent persisted before dispatch, and a fresh SQLite ledger pinned by schema digests.
- Fail-closed execution. Stale data, sequence gaps and silent feeds block decisions. Unknown order outcomes halt execution until reconciliation and are never blindly retried.
- Risk can veto, never trade. The risk guard may reject an order; it never chooses a different one.
- Labeled market data. Coinbase public REST/WebSocket (no credentials), recorded replays or synthetic fixtures, each labeled end to end.
flowchart LR
Observe --> Sense --> Propagate --> Decode --> Map --> Admit --> Execute
Execute -. research learning only .-> Reinforce
| Stage | What happens |
|---|---|
| Observe | Read an order book from a labeled source. |
| Sense | Four fixed features become an 8-bit RGB frame delivered through a photoreceptor proxy (an engineered input, not a market model). |
| Propagate | The native kernel runs the whole graph. |
| Decode | A fixed decoder reads the DNp20 (left/right) and DNpe017 descending neurons as BUY, SELL or HOLD. |
| Map | A fixed mapping moves only the chosen instrument's target weight, by at most the configured step. |
| Admit | An exact planner, turnover guard and risk actor check capital, exposure, drawdown and loss limits. |
| Execute | The intent is stored, filled in the paper book and reconciled. |
| Reinforce | Research learning only: paper profit/loss stimulates two identified dopamine cells (PAM11, PPL101), and a bounded three-factor rule adjusts existing KC→MBON synapses. |
Reinforcement is an engineered input. It does not model pain, pleasure or consciousness, and profitable learning is unverified. See the autonomous lifecycle for the full loop.
| Area | Available | Not yet supported or qualified |
|---|---|---|
| Execution | Paper (SANDBOX) spot and synthetic perpetuals with an exact local ledger | Real-money trading; live mode is closed by default |
| Market data | Coinbase public REST and WebSocket level2, JSON-lines replay, synthetic fixtures | Other exchanges |
| Controllers | Full-graph baseline inference, frozen checkpoints, single-slot paper research learning | Any validated trading result |
| Platforms | macOS ARM64, Windows AMD64 and Linux AMD64 from source | Native desktop smoke tests and native runs on every target |
| Credentials | macOS Keychain, Windows Credential Manager, Linux Secret Service; no plaintext fallback | Linux native execution-vault bridge |
| Storage | Fresh 1.0.0 SQLite ledger pinned by schema digests | Migration or import of older or modified ledgers |
Details: implementation status, venue capabilities and the latest validation evidence.
Requirements: Python ≥3.11 linked to SQLite ≥3.51.3 with Tk, uv, and CMake with a C++ toolchain for the native kernel.
python dev.py setup # create the locked environment
python dev.py prepare # download and verify the MaleCNS v1.0 dataset (about 1.1 GB on first use)
python dev.py kernel # build the native spiking kernel
python dev.py run # start the desktop appOn Windows use py dev.py … or ./dev.ps1 …; on macOS ./dev.sh … is
equivalent. In Overview, configure the malecns-autonomous strategy (capital,
slot limit, feed and controller) and enable automation. The
paper startup runbook walks through each step.
Paper mode needs no exchange credentials, and application data lives under
~/.devsolux/coinfly.
python dev.py check # lint, strict type checks and the full test suite
python dev.py test # tests only; pytest arguments pass through
python dev.py test-ui # native desktop smoke tests (needs a desktop session)
python dev.py doctor # environment and production-readiness diagnosticsCI runs lint, type checks, tests, package verification and a native compile on
macOS, Windows and Linux for every pull request and push to main. Project rules
for contributors and agents are in AGENTS.md.
macOS release builds (./dev.sh build-unsigned, ./dev.sh build, optionally with
--preflight) stay closed until genuine production prerequisites are supplied. See
desktop packaging and
macOS production builds, which also covers the
opt-in release workflows and how to verify their Sigstore signatures.
The documentation index covers operation, runbooks, venues and architecture. Start with:
- Paper startup without credentials
- Full-graph provenance and dataset bootstrap
- Experimental training and evaluation
- Safety invariants and supporting tests
Coinfly's source code is released under the MIT License. The MaleCNS v1.0 connectome data is not bundled: it is downloaded from the upstream FlyEM/Janelia release on first use and remains subject to its publisher's terms.