Phase 19-20 established the branch-isolation principle: portfolio-level constraints (gross-exposure limits) are fully disabled at the branch training level. Only per-symbol anti-pattern regulators are active, and each is gated by deadzone thresholds so it stays passive in the normal operating range and only fires on catastrophic patterns.
action_l2. Anti-saturation against tanh saturation.
action_diversity. Anti-mono-collapse — penalty for low std(actions, dim=symbols).
direction_inbalance. Anti-mono-direction — penalty when |batch_mean(actions)| exceeds threshold.
actor_cvar_aux_coeff. Hybrid actor target: mean(Q) + coeff × CVaR_α(Q). Distributional Q with per-symbol CVaR penalty favours tighter left tails.
q_crossing_coeff. Per-sym quantile monotonicity (numerical stability).
New v2.4.18 knobs.
csc_momentum / csc_eps. CrossSampleCenter running-mean EMA momentum (default 0.01) and ε for numerical stability.
csc_apply_actor / csc_apply_critic. Independent toggles for the symmetric CSC surfaces (always paired in production).
world_model_period — planner activation cadence (default every 80 cycles).