plot_skills

Paper figures from executed evidence

The application chooses figures after statistical execution. Each independent drawing request receives the complete executed result table, original source tables, research context, figure purpose/requirements, shared terminology/style, and relevant helper source.

Use the executed values and their actual scope, units, comparison direction and uncertainty. Raw source rows support actual scatter/distribution/trajectory plots; supplied summary statistics do not imply an unobserved sample. Missing cells and intervals are not zero. Do not recompute tests, create intervals, invent observations or select favorable results.

Choose a useful geometry for the research question. Native helper source is supplied and installed as global functions. Reuse those functions when they fit and call their exact signatures. Ordinary Matplotlib code may implement a required geometry with the shared appearance. Do not import skill scripts, copy example data, or rebuild interface validators.

Generate readable English axis labels, ticks, legends and annotations from the scientific meaning. Internal result IDs and raw filter expressions belong in provenance, not on axes. Use the shared terminology, series order and colors consistently across figures. Identify every displayed series with a legend or direct label; derive legend handles and labels together from the actual plotted artists. Use concise category names; define scientific abbreviations and full conditions in the external caption. State shared factor names, units and uniform sample support once. For multifactor conditions use short single-line level combinations, aligned factor columns, group headings or facets, instead of multiline paragraphs and the same n on every row. Keep group-specific sample sizes when they differ or affect interpretation. Choose this label hierarchy before wrapping; retain all conditions and comparison directions. Keep axis titles brief; put lengthy factor-order keys and metric definitions in the caption or a horizontal key instead of a long vertical title or tick label. Leave font families and rcParams to the host; emphasize with fontweight. Export uses a Unicode-capable font chain even if generated code overrides the family. Wrap long comparison names or choose a horizontal layout while retaining the full meaning. The caption explains how to read the plot, groups, units, sample/interval definitions and provenance. Numeric contrasts and substantive conclusions are written once in the separate figure analysis, rather than generated again in the drawing caption. Do not silently change scales or denominators, and omit redundant dimensionless unit suffixes.

Keep the full figure title and caption outside the image. Preserve actual negative values and interval endpoints. Differences need clear subtraction order and a zero reference. Nonnegative bars use a zero baseline. Provide sufficient space for labels and legends. Choose layout appropriate to the real data instead of copying sample margins or limits. The host applies fonts and exports SVG/PNG/PDF; return a Figure without saving or showing it.