blob: ea92ac594946e9ece3135a6b714aa02a15d32398 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
|
# RRM Report Bundle 2026-06-03
This bundle contains report/PPT-ready artifacts for the recursive reasoning dynamics work.
## Recommended Figures
- `figures/fig0_motivation_lambda1_success_failure_hrm_trm.png`: motivation, first Lyapunov exponent separates success and failure for HRM/TRM.
- `figures/fig1_hrm_trm_training_curves.png`: HRM/TRM baseline versus multi4 training curves.
- `figures/fig2_accuracy_vs_chaotic_volume_phase.png`: accuracy versus chaotic volume view.
- `figures/fig3_hrm_trm_success_failure_spectra.png`: success/failure full-spectrum separation.
- `figures/fig4_ptrm_same_subset_comparison.png`: PTRM same-subset comparison.
- `figures/fig5_qhead_vs_lambda1_ptrm.png`: PTRM Q-head score versus finite-difference stability proxy.
## Tables
- `tables/meeting_figures_v2_report.md`: concise figure strategy and caveats.
- `tables/hrm_trm_redesigned_summary.csv`: HRM/TRM headline accuracy and spectrum metrics.
- `tables/fig5_qhead_vs_lambda1_ptrm_summary.csv`: Q-head versus stability statistics.
- `tables/*eval*.csv`: eval curves and checkpoint comparisons used to generate the figures.
- `tables/*summary*.csv`: PTRM selection and Lyapunov diagnostic summaries.
## Raw Data
- `raw_npz/`: compact diagnostic arrays backing the main figures.
- Excludes model checkpoints, W&B full history dumps, and very large tangent-mode dumps.
## Scripts
- `scripts/make_meeting_artifacts_v2.py`: regenerates the main HRM/TRM/PTRM figure set.
- `scripts/make_q_lambda_scatter.py`: regenerates the Q-head versus stability figure.
## Supplemental
- `supplemental_figures/`: earlier exploratory PNGs that may be useful as backup slides.
|
