Built in the open, phase by phase.

Every bar is stamped with the moment it could actually have been known, to the millisecond. One reader enforces that on every query, so a decision evaluated at one o'clock may read only what existed at one o'clock. A leak is not something we hope to avoid. It is something the data layer will not permit.

The universe includes the instruments that died. Listing and delisting are recorded as historical facts, so a backtest cannot quietly drop the names that went to zero and keep only the winners. The graveyard is in the data, by design.

Research code and live code share the same path. There is no second implementation that drifts. What the backtester runs is what the paper loop runs, so a result and a live decision read from one contract, never two that merely resemble each other.

50 factors registered across anomalies and market microstructure: cross-sectional and time-series momentum, residual reversal, funding carry, low-volatility and low-beta anomalies, and more. Each measured for real predictive content before it is allowed to vote.

Yang-Zhang and Parkinson volatility estimators use the full open-high-low-close bar, not just closes, for a less noisy read on risk. Amihud illiquidity and the Corwin-Schultz spread estimator tell the engine what a position will actually cost to hold and to exit.

The survivors are blended into a single score per instrument, weighted by how much each factor's measured correlation with future returns is worth. What does not earn its place is shrunk toward zero.

Scores become positions through an EWMA and Ledoit-Wolf shrinkage covariance and a mean-variance optimization solved with Clarabel, with a rank and inverse-volatility fallback. A volatility-target overlay holds the book at a chosen level of risk, and a drawdown ladder and kill switch stand behind it.

Event-driven, with a signal computed at a bar's close filled at the next open, and funding modeled as the discrete cash flow it actually is. One transaction-cost authority prices every fill, in research and in paper alike, so the number a strategy earns is the number it would have paid for in the market.

The machinery that decides whether a signal is real: purged walk-forward, the Deflated and Probabilistic Sharpe ratios, Probability of Backtest Overfitting via combinatorially symmetric cross-validation, combinatorially-purged cross-validation, honest trial counts, and a must-beat-baseline gate. The full methodology lives on the performance page.

A gradient-boosted classifier trained on cost-honest triple-barrier labels, with isotonically calibrated probabilities so a stated seventy percent means roughly seventy percent. It is used only to scale conviction up or down. It never flips a signal.

A hand-rolled Gaussian hidden-Markov model reads the market's latent regime and throttles gross exposure when the regime is adverse. It runs filtered, with no lookahead, so its view at any moment uses only what was knowable then. When the market turns hostile, the engine carries less.

A loop that walks the live order book for realistic fills, places idempotent orders so a retry can never double-count, reconciles its own ledger against the venue, and runs off a single authoritative clock. State is written down, so a process killed mid-decision resumes exactly where it stopped. No real capital touches it.