All v2.4.18 architectural concepts are retained without modification — symmetric CrossSampleCenter (actor head + critic features), separate actor / critic encoders (v2.3 split), TransformerEncoder symbol mixer × 2 layers, ObsContextProjector residual, per-symbol distributional Q (32 quantiles, twin critics), world-model + CEM planner (TD-MPC), Pattern Memory cluster sampling, Phase 18.1 head (no LayerNorm, dropout 0.1), Phase 16 rollback (affine LN_out for rank stability), Phase 19-20 branch-isolation regulator calibration.
Compression summary (capacity vs architecture).
Encoder d_model — 128 → 64 (×2, capacity).
Q-network hidden — 256 → 128 (×2, capacity).
Critic ensemble hidden — 256 → 128 (×2, capacity).
World-model hidden — 512 → 256 (×2, capacity).
Replay buffer capacity — 3.6M → 1.5M (×2.4, capacity).
Symbol mixer layers — 2 → 2 (architecture, preserved).
Symbol mixer heads — 4 → 4 (architecture, preserved).
Distributional quantiles — 32 → 32 (architecture, preserved).
Critic ensemble depth — 2 → 2 (architecture, preserved).
TD-MPC optimization dim — 48 → 48 (architecture, preserved).
Total parameter reduction — ~×4 end-to-end while keeping every architectural primitive intact.