Alpha-v2.4.18 is an experimental TD3 (Twin Delayed DDPG) agent for multi-symbol crypto trading on N = 4 correlated pairs. Builds on the v2.4 branch / V_global separation with three additions: a symmetric CrossSampleCenter (CSC) anti-bias module applied on both actor and critic surfaces, per-symbol 16-quantile distributional Q, and a parallel TD-MPC world-model + CEM planner contour.
Key hypothesis. The DC component that accumulates in the residual stream of encoder / mixer pollutes per-feature batch means and forces the head to compensate through weight alignment. CSC removes that mean symmetrically on both sides — actor head input AND critic feature input — keeping ∂Q/∂a gradient flow consistent on both surfaces.