the find
− 0.6 mAn automated trading system that has never traded a real dollar — by design. Two sleeves: a crypto sleeve running Freqtrade in permanent dry-run, and a stocks sleeve of custom Python paper traders working a small universe of liquid ETFs via Alpaca. Both wrap a shared library that holds the serious machinery — position sizer, risk manager, ledger, promotion gate, meta-gate — all governed by a written mandate whose risk constants are locked policy: strategies may evolve, the constitution may not.
cross-section
− 1.3 mFive repositories forming one organism. The sleeves execute. The shared library enforces. A scout layer runs daily web sweeps and produces morning briefs and forward-state tracking. And underneath sits a knowledge graph — the system's shared brain — hand-curated, append-only, with dated experiment notes. That graph is actively written by two different AI systems: scheduled coding-agent routines on one side, research loops from a second model family on the other, both depositing findings into the same memory. Promotion from paper to live requires passing an explicit gate with a human veto standing behind it. A floor is meant to protect half of the peak value as core; the mandate itself files that as an aspiration of the model, not a verified guarantee.
inscription
− 2.0 mField note, dated 2026-07-03: The system's objective file originally carried an aggressive multiple with a year-end date attached. On this date the date was deliberately deleted — not the ambition, the deadline. The multiple became an undated North Star, on the reasoning that a target with a clock attached is an incentive to take risk the mandate exists to forbid. In September a clock went back on — a longer one, ending in 2028. The deadline is allowed to move; the mandate's risk limits are the part that is not.