External index: 📚 visualization knowledge base — cross-disciplinary plotting libs / style guides / palettes / domain ecosystems
22 hand-picked references for stress-testing whether an agent (Claude / Codex / etc.) can reproduce paper-grade style under visually complex conditions: corner plots, multi-element overlays, inset zooms, custom colorbars, log-log scaling, 3D, polar, lattice physics errorbar layouts, distributional RL, etc. Trivial single-line plots are excluded — those don't exercise style-transfer capability.
Mix: arXiv real-paper figures (✓ true top-conference aesthetic, may not ship full mpl code) + ChartMimic complex composites (✓ ship full matplotlib code). Each card explains why it's a stress test.
"Code-runnable" criterion: the figure must be reproducible by matplotlib given hypothetical numerical data — i.e. line/scatter/bar/heatmap/contour/imshow on data arrays. Figures requiring real-world JPEG photos (e.g. ImageNet samples) or raw photographic content are excluded — those aren't code-only-drawable. Astronomical/physics image data (numerical arrays displayed via imshow) is code-drawable and stays in.






















44 strong PASS + 8 partial PASS from manual audit of ~105 arXiv source bundles I downloaded earlier. These are real top-conference paper figures (NeurIPS / ICLR / Nature / PRL / ApJ / etc.), not curated benchmarks. Code is not always shipped — they're for style transfer use, not reproduction. Each thumb shows first 4 figures of the paper; click arXiv link for full PDF.