Five stages.
One continuous loop.
From a thesis the engine formed itself to a strategy that has survived 21 days of paper trading — running without pause and without instruction.
Runs on Claude Haiku 4.5. It writes signal logic only — entries and exits — and never its own backtest, so no strategy can grade its own homework. 30% of each batch explores new ground rather than refining what already worked.
The engine owns the simulation: the data window, transaction costs, liquidity-aware slippage, next-bar-open fills, position sizing, and every metric computed. A strategy plugs signals in and gets a verdict back. It cannot reclaim execution.
Each cycle is informed by what came before, but the record of proven edges is recomputed from results rather than rewritten as narrative — so a strategy that has been demoted cannot survive in memory as one that works.
Defined-risk options run on a separate harness rather than an extension of this one, because an options Sharpe and an equity Sharpe are not comparable numbers — which is what makes per-asset-class gating load-bearing rather than decorative. Structures are checked for bounded loss from their own payoff, so an unbounded position is rejected structurally rather than by remembering to add a rule. It produces no live positions. Generation is manual, nothing from it reaches the paper book, and its thresholds were set by measuring what noise scores — a test the first attempt did not pass.
What it has produced.
Every strategy the engine has written, grouped by the thesis behind it and the family it belongs to. Almost all of it is rejected — that is the mechanism working, not failing.
The gates themselves — every threshold, and why each one exists — are set out in full on the method page.