make-model-figures

Make Model Figures

For paper-style refinement, read references/cumcm-research-style.md: use the shared render-time configuration, chart-specific layout, and an individual three-case review for each affected entry. Keep visual decisions separate from numerical and export checks.

Build a visual argument, not a gallery of available variables. Every formal figure must support one modeling claim and remain truthful to the saved data, model, units, and uncertainty.

Choose the operating mode

• Create: design a new figure from machine-readable data or saved model results.
• Improve: preserve the source data, statistical meaning, units, filtering, and conclusion while revising layout, chart type, typography, palette, or annotations. Record any changed scale, aggregation, normalization, or encoding.
• Audit: inspect the source and rendered output; report evidence or readability defects without silently redrawing it.
• Explore: make lightweight diagnostic plots. Do not represent them as submission-ready until the formal contract and export checks are completed.

Proceed directly when the user's goal and data mapping are clear. Ask only when a missing choice would change the scientific meaning, such as which columns represent the requested comparison.

Required workflow

1. State the visual claim. Read references/paper-figure-language.md and write the data-derived conclusion (or an explicit pending question with claim_status="question"), figure role, evidence, source paths, units, sample definition, statistic, uncertainty, model/scenario, and review risks. Use FigureContract for a formal figure; use a shortened claim/evidence note for exploration or a cosmetic-only edit.
2. Protect data integrity. Use all relevant user-provided observations. Never silently select representative rows, replace real data with illustrative values, or drop inconvenient groups. For dense rendering, prefer rasterized marks, hexbin/density views, or transparent points; if aggregation, filtering, imputation, winsorization, or normalization changes what is shown, record it and keep the unmodified source.
3. Route by information task. Read references/figure-standards.md. When the chart type is unclear, specialized, or multi-panel, also read references/design-and-routing.md. Prefer the smallest set of non-redundant panels that proves the claim; avoid duplicate views of the same fact.
4. Design before coding. Fix the final physical width and approximate aspect ratio first. For two or more panels, create a FigureDesignBrief: core message, layout recipe, hero panel where appropriate, support sequence, and stable color roles. A hero panel should be larger only when its evidence is genuinely more important or denser.
5. Assign semantic style. For new figures or recoloring, read references/palette-system.md. Choose categorical colors and continuous maps separately by object and physical quantity; keep the same object's meaning stable across figures. Preserve useful visual emphasis in a supplied original. For power/energy maps or markers over a heatmap, also read references/power-energy-figures.md. Retain line/marker/label redundancy for critical distinctions. Do not change data or statistics during recoloring.
6. Generate reproducibly. Prefer algorithms.figures.get_figure/get_spec/list_figures for registered charts; algorithms.sci_figures remains the compatibility and export entry point. Plot from saved data/results with the registered renderer or algorithms.diagram when available; otherwise install the repository-built math-model-agent wheel; use the verified scripts/minimal_example.py to diagnose the runtime. Do not copy an algorithm implementation into the skill. Read references/execution-and-migration.md for configuration precedence, six executable routes, installation and schema migration. Use load_figure_config and render_template for common tasks. Apply publication_rc_params or paper_figure_rc_params, exact publication_size, consistent add_panel_label, and export_publication_figure. Do not assume the project contains a code/algorithms checkout.
7. Run the quality gate. For a formal figure, read references/publication-workflow.md and references/visual-qa-checklist.md. Run the programmatic audit, export the vector master and 450 DPI proof, then actually open the proof at final paper size. Fix clipping, overlaps, missing glyphs, misleading scales, unstable color meanings, weak hierarchy, and illegible details before delivery.
8. Close the evidence loop. Introduce the figure in the manuscript, explain its pattern and modeling meaning after it, and state limitations. Confirm every plotted optimum, ranking, threshold, sample size, interval, and annotation against the saved result and surrounding tables/text.