| Age | Commit message (Collapse) | Author |
|---|
|
warning; lighter FTLE default
- extended_rollout: per-row drift norm was a scalar -> np.stack shape mismatch crash. fixed + verified.
- setup: warn + Colab GPU instructions when no CUDA (CPU is very slow for the JVP cell).
- FTLE default n=128->64 (still AUC~0.99) for faster runs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MZBxRQ65wDxiUSm9Hr5Ere
|
|
add progress bars
- NEW cell 2: leading FTLE (Benettin/JVP along trajectory) -> success-vs-failure separation FIRST.
Validated vs paper: TRM median l1 +0.01(succ)/+0.11(fail) AUC 0.99; HRM -0.15/+0.02 AUC 1.00.
- remove the toy-model section (lived as a distraction before the real result)
- tqdm progress bars on FTLE / rollout / basin cells
- reorder: load -> lambda_1 result -> transient-chaos escape -> basin accessibility
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MZBxRQ65wDxiUSm9Hr5Ere
|
|
framing + basin cell
- point loader at the uploaded HF model bundle (TRM/HRM Sudoku checkpoints + code + test set)
- correct mechanism text: both arches are chaotic SADDLES (transient chaos), not attractors;
escape-rate gap is mostly compute-per-segment (~1.6x per module-eval, not 5x)
- add cell 4: z0-perturbation basin-accessibility (input- vs IC-determined trapping)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MZBxRQ65wDxiUSm9Hr5Ere
|
|
Notebook: toy transient-chaos (no GPU) + load TRM/HRM from HF + extended rollout
(TRM escapes / HRM trapped). Model loading verified locally (missing=0 both).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
