# pr-figure — autoresearch journey figures (data + scripts + code)

Self-contained package for the autoresearch writeup figures.

```
pr-figure/
  data/      all run data (json/tsv/csv)  — 795 H100 runs + 59 B200 runs + the BO/wave results
  scripts/   plot_*.py  (portable ones read ../data, write ../figures)
  figures/   rendered PNGs
  code/      source files behind the code snippets (opt_param_full2.py, bo_champ.py)
  JOURNEY_CODE_SNIPPETS.md   each plot rung (+ Muon / value-emb / n-gram / softcap / GP …) -> verbatim code
  BUNDLE_README.md           figure -> script -> data map + provenance
```

## Regenerate the figures (portable scripts)
```bash
python3 scripts/plot_journey_dense.py     # figures/progress_journey_dense.png
python3 scripts/plot_progress_messy.py    # figures/progress_messy.png
python3 scripts/plot_b200_from_csv.py     # figures/b200_iteration.png
```

## What's where
- **data/all_remote_val_bpb.json** — every H100 run's val_bpb (the master dataset, 795 runs).
- **JOURNEY_CODE_SNIPPETS.md** — the labeled rungs on the journey plot, each mapped to the real code
  (verbatim, with `file:line` anchors into `code/`).
- **BUNDLE_README.md** — full figure→script→data table + the note that `front_densify_pts.json` was
  reconstructed from the remote logs (the original was lost from /tmp).