Data-figure re-rendering from frozen tables

Iron rule

Every point on a data figure is real experimental data — the encoding is the semantics. Faithful replication = re-render from the same frozen tables (CSVs), keeping palette, marker shapes, axes, log ticks, legends and count annotations item-by-item. Never hand-draw, restyle, or move points — that is fabrication, not replication. Legitimate upgrades are exactly three: an SVG vector version, a higher-dpi PNG, and fixing objective defects in the original (text overflow, label colliding with a marker, edge clipping).

Workflow

1. Locate the real data. A paper's figure script often carries a stale relative path — find <project-root> -name "<key>.csv" for the live outputs/tables directory; verify every CSV exists with plausible row counts before writing code.
2. Port the data pipeline verbatim. Copy the original script's load() / aggregation / pairing functions line-for-line; never change a statistic.
3. Upgrade rendering. save(fig, name) emits both SVG (svg.fonttype="path" — text as paths, renders identically everywhere) and PNG (600 dpi). One script drives all figures: python make_figures_svg.py regenerates everything.
4. Acceptance. Vision-judge against the original PNG: panel structure, axes, legends, annotations identical + point clouds coincide (same source data ⇒ points coincide naturally, which also cross-validates the ported pipeline).

Template

make_figures_svg.py in this folder is a complete runnable template (six figure types from one script): global rcParams, dual-format save(), and ready-made patterns for overview block diagram / dot-strip panel / paired log-log scatter / slope chart / grouped bars / histogram. Adapt load() to your own frozen tables.

Read before drawing

PITFALLS.md — ten field-tested pitfalls, including: no linear text offsets on log axes (raster overflow at 600 dpi — use multiplicative), FancyBboxPatch clips to the axes bbox (anchored-at-edge boxes silently lose their bottom border), multi-line text anchor semantics, box-title width overflow, stale relative data paths.