Architecture

Actor pathway.
  obs → actor_encoder (SnapshotEncoder, d_model = 64)
      → actor_symbol_mixer (TransformerEncoder × 2, 4 heads)
      → ObsContextProjector (residual context)
      → actor_mean_head [Linear → GELU → Dropout(0.1) → Linear]
      → clamp(±3) → tanh → action ∈ [−1, +1]

Critic pathway (twin Q-ensemble, no CSC).
  obs → critic_encoder (separate from actor — v2.3 split)
      → critic_symbol_mixer (TransformerEncoder × 2, 4 heads)
      → ObsContextProjector
      → concat(action) → DiverseCritic × 2 (32-quantile distributional Q)

Key architectural difference vs v2.4.18-mini. No CrossSampleCenter modules — neither on the actor head nor in the critic feature pathway. Topology is otherwise structurally identical.

Compression summary (matches v2.4.18-mini for apples-to-apples comparison).

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 — 1.2–3.6M → 1.5M (aligned with v2.4.18-mini).

Symbol mixer layers — 2 → 2 (architecture, preserved).

Symbol mixer heads — 4 → 4 (architecture, preserved).

Distributional quantiles — 32 → 32 (architecture, preserved).

TD-MPC optimization dim — 48 → 48 (architecture, preserved).

Preserved baseline elements. Per-symbol distributional Q (32 quantiles, twin critics), separate actor / critic encoders (v2.3 split), ObsContextProjector residual, 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), Phase 19-20 branch-isolation regulator calibration.