TD3 with deterministic policy, target-policy noise smoothing, and policy_delay = 1.
Update schedule. 2 critic steps and 2 actor steps per outer iteration. Twin critics polyak-targeted alongside encoders, mixers, and CSC running means.
Return target. n-step TD return with n = 4 and γ = 0.985. Distributional Huber on per-sym quantiles, REDQ-style min between paired ensemble critics.
Replay. Buffer capacity 3.6M transitions. Batch size 12 000 (RTX 5090-class compute). Pattern-memory sign-balanced sampling — UP / DOWN / FLAT clusters with fixed quotas.
Symbols. N = 4 correlated crypto pairs.
World-model planner. CEM lookahead activated every 80 cycles; runs alongside the TD3 contour but does not steer the gradient — it provides verification and tie-breaking at difficult decision points.