| # Report | |
| Comprehensive write-up of the LeWM × PushT control experiment: the | |
| horizon-reset procrastination pathology, the arrival-and-hold fix, and the | |
| ablation study. | |
| ## Contents | |
| | path | what | | |
| |---|---| | |
| | `report.tex` | the document — abstract, full formula inventory, ablations, statistics | | |
| | `figures/` | eight figures, each as `.pdf` (for LaTeX) and `.png` (for viewing) | | |
| | `tables/` | five LaTeX tables, `\input` by the document | | |
| | `make_figures.py` | regenerates `figures/` from the raw result files | | |
| | `make_tables.py` | regenerates `tables/`; also recomputes and persists the paired statistics | | |
| | `recover_profiles.py` | recovers the original controller's per-offset validation profile (its checkpoint predates that logging) | | |
| | `check_tex.py` | static check of the `.tex` — dangling refs, missing inputs/figures, brace and math balance | | |
| | `build.sh` | regenerate everything, then compile | | |
| ## Building | |
| ```bash | |
| bash report/build.sh | |
| ``` | |
| Or step by step: | |
| ```bash | |
| python report/make_figures.py # -> report/figures/ | |
| python report/make_tables.py # -> report/tables/ | |
| latexmk -pdf report/report.tex # -> report/report.pdf | |
| ``` | |
| **No LaTeX engine is installed on the machine this was written on**, so the | |
| PDF has not been compiled here — `report.tex` passes `check_tex.py` but has | |
| not been through a real typesetter. To build it, install one of: | |
| - Windows — `winget install MiKTeX.MiKTeX` | |
| - macOS — `brew install --cask mactex-no-gui` | |
| - Debian/Ubuntu — `sudo apt install texlive-latex-recommended texlive-fonts-recommended` | |
| - Anywhere — `cargo install tectonic` (single binary, fetches packages on demand) | |
| Or upload this directory to Overleaf; it is self-contained and uses only | |
| stock packages (`amsmath`, `booktabs`, `graphicx`, `caption`, `subcaption`, | |
| `xcolor`, `microtype`, `hyperref`, `geometry`). | |
| ## Provenance | |
| Every number in the document is generated, not transcribed. The scripts read | |
| only: | |
| - `data/runs/eval/results.jsonl` — closed-loop evaluations | |
| - `data/runs/diagnostics/` — contraction fits, refinement traces, support statistics | |
| - `data/runs/*/controller.pt` — checkpoints, for hyperparameters and validation profiles | |
| so the prose cannot silently drift from the results. The paired statistics | |
| are recomputed by `make_tables.py` using the same functions as | |
| `scripts/paired_stats.py` and written back to | |
| `data/runs/eval/paired_stats.jsonl`. | |
| One figure input is not a result file: `fig_training_signal` parses raw | |
| training logs, two of which live under the job temp directory. If those are | |
| gone, that figure will render with fewer series; everything else is | |
| unaffected. | |