paper-data-figurelisted

# Paper data figure The sibling skill draws what has no data behind it. This one draws what does, and the whole difference in outcome comes from two habits: the numbers are read from the files the experiments wrote, and the figure is drawn at the size it will be printed at. ## Hard rules 1. **Draw at final print size; `\includegraphics` never scales.** Set `figsize` to the true width in inches and include the file with no `width=` at all. `width=0.48\textwidth` on a figure drawn at default size is how an 8 pt label silently becomes 5 pt, and nothing in the compiled PDF announces it. 2. **A LaTeX point is 1/72.27 inch; a PostScript point is 1/72 inch.** `\the\textwidth` reports LaTeX points. Convert through inches. Treating the two as the same number puts a CVPR full-width figure 0.4 % over the text width, which is enough for `Overfull \hbox`. Measure the venue once, in the document body: `\typeout{\the\textwidth, \the\columnwidth}`. 3. **Numbers come from the result files, never from a message or a memory.** Read the JSONL, CSV or summary the run produced. A figure that hardcodes a number is a figure that will disagree with its own table after the next run. 4. **Never plot an arm that has not run.** An empty series is left out and said so in the caption; it is not interpolated, extrapolated, or filled with a plausible value. 5. **One script per figure, one shared style module, imported not copied.** `scripts/paper_plot_style.py` holds the geom