... into a backtest. The backtest produces data. That data, together with earlier conversations, becomes context for the next LLM call. The next thought has a foundation. I want each call to leave something in these layers: trading algorithms as granite, factors as basalt, machine learning models as obsidian. Work that fails validation leaves a record too, for the next call to consult. The LLM is a visitor passing through. The code, the data, and the history stay, slowly becoming ...