Live experiment: does weekly human review make AI trading decisions better over time? Week 2 update

Been running this since March 17. Sharing week 2 data. Setup: Multiple AI models analyze live crypto market conditions hourly and make decisions on a $1,000 paper portfolio. Every Sunday I review all outcomes and adjust the rules. No backtests — everything runs against live market data. The research question: does the iterative human review loop measurably improve AI decision quality over 26 weeks? Week 2 numbers: – Portfolio: -0.3% vs BTC -5.8% — preserved capital during the drawdown – Signal accuracy (4h): 78% – When all models agree: 85% accuracy – When only one model fires: 25% — this is the main problem to solve Tracking every decision at 4h, 24h, and 48h intervals. Live at theaitradinglab.com Early days but the data is accumulating. Happy to discuss methodology.