rrm workspace: TRM/HRM/SRM code, Maze dataset, dynamical-analysis pipelineHEADmain

Curated export for clone-and-run Maze training (2x A6000) + diagnostics.
trm/hrm pretrain.py carry trajectory-augmentation code (backward-compatible).
Heavy artifacts (checkpoints/wandb/npz) gitignored; see PROVENANCE.md.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2 files changed, 42 insertions, 0 deletions

diff --git a/srm/config/arch/hrm_v1.yaml b/srm/config/arch/hrm_v1.yaml
new file mode 100644
index 0000000..a5646b8
--- /dev/null
+++ b/srm/config/arch/hrm_v1.yaml
@@ -0,0 +1,21 @@
+name: hrm.hrm_act_v1@HierarchicalReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 2
+L_cycles: 2
+
+H_layers: 4
+L_layers: 4
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
diff --git a/srm/config/arch/srm_v1.yaml b/srm/config/arch/srm_v1.yaml
new file mode 100644
index 0000000..1b5ece5
--- /dev/null
+++ b/srm/config/arch/srm_v1.yaml
@@ -0,0 +1,21 @@
+name: srm.srm_aol_v1@StableRecursionModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+# SRM-specific
+n_iters: 12           # joint micro-steps per ACT step (≈ HRM's H_cycles·L_cycles+H_cycles = 6 with deeper schedule)
+n_aol_layers: 2       # depth of AOL ψ block (channel + token mix per layer)
+kappa: 0.9            # contraction factor: per-step Lip_P ≤ (1-α)+α·κ = κ
+eta: 1.0              # weighting of L block in P-norm (1.0 = symmetric)
+alpha: 1.0            # damping (1.0 = full step)
+
+hidden_size: 512
+puzzle_emb_ndim: ${.hidden_size}
+
+# Unused (kept so pretrain.py's __pydantic_extra__ doesn't break)
+# pretrain.py's create_model() passes some fields HRM expects; Pydantic 'ignore'
+# (default) drops them silently.