Regulator set identical to v2.4.18 — Phase 19-20 branch-isolation principle, deadzone-gated anti-pattern penalties (action_l2, action_diversity, direction_inbalance), distributional CVaR penalty in the actor objective, q_crossing for per-sym quantile monotonicity. CSC knobs (csc_momentum, csc_eps, csc_apply_*) carry over unchanged. No regulator-level differences between full and mini — only the underlying capacity scales.
Role in the experimental matrix. v2.4.18-mini occupies the lower-right cell of a 2×2 grid spanning {capacity: full / mini} × {CSC: off / on}:
full d_model 128 : v2.4.4 v2.4.18 mini d_model 64 : v2.4.4-mini v2.4.18-mini
Horizontal pairs isolate the CSC effect at each capacity scale; vertical pairs isolate the capacity effect at each architecture; the diagonal exposes interaction effects. Without the complete 2×2, any conclusion about CSC vs capacity stays confounded.
Trade-offs. Pure size reduction preserves full architectural semantics at the cost of: lower representation headroom (though full-model capacity utilization was already ~14% of d_model = 128), slower convergence to peak performance on large datasets, reduced buffer diversity span. v2.4.18-mini is a research tool for controlled comparisons and methodology validation — not a deployment target.