Internal production system · selective access by agreement
One strategy contract across backtest, paper, sandbox and live environments, with deterministic replay, attributed execution, central risk, venue adapters and reproducible research workflows.
Measured engineering surface
Repository counts come from the current private codebase. Performance figures are explicitly scoped to their benchmark rather than presented as trading results.
Feed parsing and aggregation benchmark—not exchange latency or investment performance.
| Measure | Count | Signal |
|---|---|---|
| Named test functions | 1,706 | |
| Python test files | 426 | |
| Contract-test modules | 47 | |
| Execution modes | 4 |
Counts are generated from the repository tree; bars use a logarithmic scale and do not represent coverage percentage.
Architecture
The runtime keeps decision provenance intact while separating shared state, strategy logic, controls and venue-specific execution.
Product
Strategies target stable contracts. The runtime owns shared state, attribution, risk and inventory; execution adapters handle what must differ between deterministic simulation and real venues.
One session coordinates reference data, market state, strategies, routes, inventory, risk and control-plane services.
Market inputs produce one immutable snapshot per step, then filtered strategy views preserve isolation without duplicating state.
Strategy, execution and route identities travel with intents, reports and diagnostics instead of disappearing inside broker code.
Risk decisions, execution adapters, post-trade sinks and kill-switch controls remain first-class boundaries rather than strategy side effects.
Execution modes
Historical replay with a strong deterministic expectation for research, regression testing and reproducibility.
Live or replayed inputs with simulated execution for strategy shakeout without venue-side orders.
Real broker adapters against testnet or venue sandboxes for integration and pre-live validation.
Real market input and venue execution, with operational safeguards and environment-dependent timing.
The strategy-facing contract remains stable; fill formation, clocks, venue capability and safety defaults are allowed to differ explicitly.
Deterministic replay
Runtime history, snapshots, attribution, order transitions and versioned checkpoints support replay, recovery and diagnosis without collapsing everything into an opaque backtest result.
Read the runtime guaranteesVenue boundaries
Venue details stay behind feed, broker and execution-adapter boundaries, while reference data normalises instruments and constraints for the rest of the runtime.
CCXT-backed market data and execution, with REST, streaming and historical ingestion paths.
Dedicated Polymarket and Kalshi sessions, feeds, stores, brokers and simulation surfaces.
Uniswap and PancakeSwap connector architecture with quoting, slippage, token and transaction boundaries.
Derived OHLCV can be built from trades, top-of-book observations and pool-state snapshots for a consistent research surface.
Research discipline
Research workflows around the runtime preserve the request, strategy, dataset and configuration identity behind each result.
Walk-forward windows, purged cross-validation and embargo periods for time-series evaluation.
Deterministic request, strategy, dataset and configuration hashes with versioned metadata.
Reports, trades, equity curves, stress reruns and failure capsules stored per evaluation.
Queryable run registry, report comparison and leaderboard tooling without publishing proprietary results.
Private performance · public infrastructure
Trading results, strategy logic and live data are proprietary. Public evidence focuses on the runtime surfaces that can be described without exposing edge: execution modes, replay, risk boundaries, venue adapters, research validation, reporting and test coverage.
More than 400 Python files in the test tree, spanning unit, contract, integration, end-to-end, performance and parity concerns.
Explicit timestamp, snapshot, attribution and order-lifecycle guarantees protect replay and regression behaviour.
Intentional differences between backtest, paper, sandbox and live behaviour are recorded instead of silently leaking across modes.
Benchmark and profiling infrastructure covers data processing, ordering, memory and runtime paths, with synthetic results identified as such.
Selective access
CrackTrader is an internally used product, not a public download, hosted trading service, signals product or managed-investment offering. Access to the software, architecture, private demonstrations or deployment support may be considered case-by-case for serious technical, commercial or research enquiries.
Potential discussions may include licensing, collaboration, deployment support or deeper technical review. Strategies, performance data and proprietary datasets remain private unless separately agreed.