97a770da65
Flight7 model underestimates by 80% in CWM dynamic book: raw predicted: 0.034 bps, actual: 0.180 bps Constant error across 22K episodes — no feedback loop. Root cause: Flight7 calibrated on real BingX taker fills, but CWM's synthetic dynamic book has different fill characteristics. Fix: SlippageSelfCalibrator with EWMA feedback loop. After each fill: error = actual - predicted (clipped to +/-20 bps) EWMA smooths per-symbol errors (alpha=0.2) Next prediction = raw_model + EWMA_correction Bounded output: 0-50 bps absolute Convergence (300 eps across 8 assets): ETH: 9% error (from 80%) SOL: 3.5% DOGE: 6.7% LINK: 5.7% ADA: 9.7% BTC: 48.6% (low fill count, converging) AVAX: 28.6% (low fill count) UNI: 52.5% (low fill count, early outlier) Truthfulness guarantees: - Correction is observable (CALIBRATOR.correction(symbol)) - Resets between runs (no hidden state) - Only uses observed fills, no assumptions - Error clipping prevents outlier domination - Absolute bounds prevent runaway