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, 43 insertions, 0 deletions

diff --git a/research/flossing/flossing_suite/results/toy_rnn/toy_prefloss_N80_k40_E1000.cmd.sh b/research/flossing/flossing_suite/results/toy_rnn/toy_prefloss_N80_k40_E1000.cmd.sh
new file mode 100755
index 0000000..db82a74
--- /dev/null
+++ b/research/flossing/flossing_suite/results/toy_rnn/toy_prefloss_N80_k40_E1000.cmd.sh
@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "/home/yurenh2/rrm"
+export CUDA_VISIBLE_DEVICES="1"
+export PYTHONUNBUFFERED=1
+exec "/home/yurenh2/miniconda3/envs/rrm/bin/python" research/flossing/engelken_python_flossing.py \
+  --hidden-size 80 \
+  --n-lyap 40 \
+  --train-epochs 1000 \
+  --inter-period 100 \
+  --inter-epochs 100 \
+  --batch-size 16 \
+  --input-dim 1 \
+  --train-steps 300 \
+  --lyap-steps 55 \
+  --floss-input-steps 300 \
+  --seed-ic 1 \
+  --seed-input 1 \
+  --seed-net 1 \
+  --seed-ons 1 \
+  --lr 0.001 \
+  --beta1 0.9 \
+  --beta2 0.999 \
+  --init-type 1 \
+  --recurrent-gain 1.0 \
+  --recurrent-mean-gain 0.0 \
+  --input-scale 1.0 \
+  --delay 10 \
+  --ws-std 1.0 \
+  --ws-mean 0.0 \
+  --wr-std 1.0 \
+  --wr-mean 0.0 \
+  --b-std 0.1 \
+  --b-mean 0.0 \
+  --task -1 \
+  --lyap-target 0.0 \
+  --eval-every 100 \
+  --eval-batches 4 \
+  --log-every-floss 25 \
+  --device cuda \
+  --pre-epochs "100" \
+  --max-inter-episodes "0" \
+  --out "/home/yurenh2/rrm/research/flossing/flossing_suite/results/toy_rnn/toy_prefloss_N80_k40_E1000.json"