jgrusewski
a5945db538
spec: Adaptive Training Dynamics v2 — 4-layer design for breaking Sharpe plateau
Layer 1: Rank-preserving signed reward standardization (SNR 0.001→0.5)
Layer 2: Temporal Q-gap momentum (spread modulation at plateaus)
Layer 3: Direction-conditioned magnitude (dir Q-values → mag head input)
Layer 4: Distributional variance position sizing (Kelly from C51 atoms)
Based on train-w6qfd results: val_Sharpe=21-25 sustained, Q-gap=0.138
growing. Builds on 32 fixes from this session. Implementation order:
1→2→4→3 (signal → momentum → variance → architecture).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>