This is a systematic trading research lab with the machinery you'd expect at a fund: five pods — trend, cross-sectional momentum, volatility premium, FX carry, and an incubator — each with a written mandate, a risk budget, and pre-committed kill criteria. It runs on 14.3 million rows of survivorship-aware price history covering 3,729 symbols including delisted names, plus sixty years of factor data, and it has executed 599 logged backtests across 64 strategy hypotheses drawn from the academic literature.
AI agents do the heavy lifting — implementing strategies, running test batteries, drafting research readouts across parallel git worktrees — inside a pipeline whose rules are enforced by code: rules frozen in writing before any formal test, at most four free parameters per strategy, and one holdout touch per hypothesis, ever.
Statistics you can actually trust
What makes the lab institutional-grade is the multiple-testing discipline. Every backtest ever run gets a row in an append-only registry, and deflated Sharpe ratios are computed against the full trial count — the correction most retail quants (and plenty of professionals) skip. The pipeline proved itself by catching exactly what it was designed to catch: a strategy that posted a 5.85 Sharpe in training collapsed to −0.31 in frozen validation, flagged and killed before a dollar of risk. When a result survives this gauntlet, it means something.
The survivor
One has: developed-markets profitability (RMW) held up through train, validation, and a holdout it was allowed to touch exactly once — Sharpe 0.88 → 0.77 → 0.50, with a −12.6% max drawdown and negative correlation to developed-market beta. It's now promoted for implementation research. A 1-in-64 survival rate is the honest base rate of edge with public data — and the entire point of the machine is that its one survivor is real.