An agentic research and execution system. Nine stages, each an independently runnable script that writes a file the next one reads — so a crash never loses the chain and every step stays auditable afterward. What follows is the design, not the deployment: the production system runs privately, and nothing here shows holdings, performance, or account data.
The governing rule
The model proposes. Code disposes.
The language model does judgment — research, scoring, thesis. Deterministic scripts do gatekeeping and measurement. A code-level DENY is unconditional: there is no override and no retry with better phrasing. That split is the entire point, because it is what stops a model allowed to touch real money from talking itself into a position.
Why it is built this way. Trades are deliberately rare, which means real outcomes arrive far too slowly to learn from. So the system manufactures its own sample: every scored candidate is also carried as a virtual trade across several exit policies, and results are measured in R-units rather than win rate — expectancy, profit factor and bootstrapped confidence intervals, so that a lucky streak cannot read as skill. A drawdown ladder steps the whole system from active, to reducing, to halted, and the validator enforces it whether or not the model agrees.