dev-viz

Set font family and size once in a shared style module (such as a publication.mplstyle file) that every figure script imports, and set figure.figsize to the target column width in inches. Avoid per-script rcParams overrides, which produce inconsistent fonts across figures.

Plot the paired per-seed difference and its interval rather than two overlapping mean±std bands. Overlapping bars are not a valid visual test in either direction, so a figure supporting a significance claim should show the difference the claim is about.

Every error bar needs the number of seeds (n) behind it and a statement of whether the interval is SD, SEM, or a bootstrap CI. These quantities differ by large factors and are indistinguishable by eye, so an unlabelled bar is unreadable evidence.

The usual cause is scaling figures with \includegraphics[width=...] in LaTeX, which rescales the text inside each figure differently. Set figsize in inches to the final column or text width in the plotting script and export at that size.

No. The generator should emit a bare tabular environment with only the rows, while the human-owned source keeps the float environment, placement, caption, and label. Emitting the whole float takes caption wording away from the author on every regeneration.

Bold should mark the pre-registered winner, not the largest number observed after the fact. Emphasis chosen after seeing the cells is an invisible selection; either drop it or state what was selected and over how many cells.