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, 36 insertions, 0 deletions

diff --git a/research/flossing/run_step4_pipeline.sh b/research/flossing/run_step4_pipeline.sh
new file mode 100755
index 0000000..dc2d86d
--- /dev/null
+++ b/research/flossing/run_step4_pipeline.sh
@@ -0,0 +1,36 @@
+#!/bin/bash
+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
+
+# Wait for Phase 1 (PID $1) to complete
+P1_PID=$1
+echo "[$(date '+%H:%M:%S')] runner waiting for Phase 1 PID $P1_PID..." >> step4_runner.log
+while kill -0 "$P1_PID" 2>/dev/null; do
+  sleep 30
+done
+echo "[$(date '+%H:%M:%S')] Phase 1 done. Starting Phase 2: CF λ*=0" >> step4_runner.log
+
+# Phase 2: CF λ*=0
+python step4_from_scratch.py \
+  --n-steps 3000 --batch-size 8 \
+  --alpha-rf 10.0 --lambda-star 0.0 \
+  --warmup-steps 200 \
+  --eval-every 300 --eval-n 512 --eval-batch-size 32 \
+  --out step4_Y_cf_lstar0_fromscratch.json \
+  > step4_Y.log 2>&1
+
+echo "[$(date '+%H:%M:%S')] Phase 2 done. Starting Phase 3: CF λ*=-0.15" >> step4_runner.log
+
+# Phase 3: CF λ*=-0.15 (natural attractor enforced)
+python step4_from_scratch.py \
+  --n-steps 3000 --batch-size 8 \
+  --alpha-rf 10.0 --lambda-star -0.15 \
+  --warmup-steps 200 \
+  --eval-every 300 --eval-n 512 --eval-batch-size 32 \
+  --out step4_Z_cf_lstar_neg15_fromscratch.json \
+  > step4_Z.log 2>&1
+
+echo "[$(date '+%H:%M:%S')] Pipeline complete (X + Y + Z all done)" >> step4_runner.log