Edge discovered by machines, validated by rigour.
Traditional quantitative research is bottlenecked by human throughput. Arctix operates differently — the engine forms its own theses, builds strategies to express them, and stress-tests each one against a fixed statistical bar it cannot argue with.
Of 20,612 strategies evaluated since 2026-04-15, 17,130 were rejected outright and 18 are still standing. That ratio is the mandate. No discretion. No override. No fatigue.
The engine produces novel strategies around the clock — exploring signal combinations, timeframes, and market regimes without human prompting. 30% of every batch is reserved for exploration rather than refinement of what already worked.
A candidate clears Sharpe ≥ 1.2, max drawdown ≤ 20%, an out-of-sample holdout, a benchmark-alpha test against SPY, and regime robustness — then trades on paper for a minimum of 21 days before it counts as validated.
The system learns from each cycle. Proven edges are recomputed from results rather than remembered as narrative, so a strategy that stops working cannot linger in the engine's memory as one that does.