Alpha-v2.1 is a refinement of v2 (TD3 + distributional quantile-regression critic on a shared cross-symbol encoder) that addresses three observed degradations of the previous architecture: latent rank-collapse in the readout adapter, bias accumulation breaking Q-head calibration, and a state-agnostic actor head emerging late in training.
Result. Composite test idx improves from −0.265 to −0.213 (+0.052, ~19% relative). Strongest gains on periods [13], [15], [17], [18]; period [19] (sideways regime) remains an architecture-independent challenge.