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>

1 files changed, 27 insertions, 0 deletions

diff --git a/research/flossing/run_HRM256_after_H.sh b/research/flossing/run_HRM256_after_H.sh
new file mode 100755
index 0000000..e901c30
--- /dev/null
+++ b/research/flossing/run_HRM256_after_H.sh
@@ -0,0 +1,27 @@
+#!/bin/bash
+# Wait for H-runner (Engelken on converged HRM) to finish, then run HRM hidden=256 baseline
+# as capacity-matched control for SRM v1 (hidden=256/512 plateau diagnosis).
+set -e
+cd /home/yurenh2/rrm/research/flossing
+source /home/yurenh2/miniconda3/etc/profile.d/conda.sh
+conda activate rrm
+export CUDA_VISIBLE_DEVICES=2
+
+H_RUNNER_PID=$1
+echo "[$(date '+%H:%M:%S')] HRM256-runner waiting for H-runner PID $H_RUNNER_PID..." >> step4_runner.log
+while kill -0 "$H_RUNNER_PID" 2>/dev/null; do
+  sleep 30
+done
+echo "[$(date '+%H:%M:%S')] H done. Starting HRM hidden=256 baseline (capacity-matched control for SRM v1)" >> step4_runner.log
+
+# HRM hidden=256 from scratch, 3000 steps, no CF (baseline for capacity diagnosis)
+python step4_from_scratch.py \
+  --n-steps 3000 --batch-size 8 \
+  --hidden-size 256 --num-heads 4 \
+  --alpha-rf 0.0 \
+  --warmup-steps 200 \
+  --eval-every 300 --eval-n 512 --eval-batch-size 32 \
+  --out step4_HRM256_baseline_fromscratch.json \
+  > step4_HRM256.log 2>&1
+
+echo "[$(date '+%H:%M:%S')] HRM hidden=256 baseline complete" >> step4_runner.log