Icarus Figures

Turn a dataset + a claim into a publication-grade figure — one a top-journal editor accepts with no revision request. Default stack: matplotlib/seaborn for data figures, original TikZ for hero/method figures, both reproducible (script + data file), exported as PDF + PNG + SVG.

First read

Open references/figure-cookbook.md and work in this order:

1. §0b (the quality bar) — the four axes (Depth · Elegance · Unimpeachable · Visible gap). Internalize this before plotting; it is the acceptance test, not an afterthought.
2. §0a (Figure Contract) — write the one-line core_claim + hero panel + what stats go on the figure, before any code.
3. §0 style preset — from _style import paper_style, save (the preset lives in scripts/_style.py). Never hand-edit rcParams per figure.
4. §A archetypes A1–A13 — pick the matching template (time series+CI, sorted bar, grouped bar, residual, heatmap, scatter+fit, Pareto, tornado, confusion, phase, network), paste, swap the data path + labels, run.

For a hero / method / framework figure (the one that carries the paper)

Read references/framework-figures.md — the deep-dive, with five compilable heroes to paste from (P1 pipeline, P2 paradigm-swimlanes, + T4 graphical-model / data-tensor / matrix-fit), the borrowable TikZ techniques, and curated journal-grade resources.

The rule to internalize first: never a generic boxes-and-arrows flowchart — embed the real method object (a distribution, a before/after scatter, a decision region), not a text label. Route through the cookbook: §I composition paradigms P1–P6 → §J craft spec → §K original standalone TikZ (compiled to PDF, \includegraphics, zero compile risk). framework-figures.md walks all three with worked examples.

Hard rules (non-negotiable)

• Reproducible only: every figure is generated by a committed script that reads data from a file path. No inline data > 20 rows. No AI-generated images (unreproducible).
• Three-format vector: save() writes PDF + PNG + SVG (text stays editable in SVG/PDF). Never ship a 72-dpi PNG.
• Style preset, once: call paper_style(...) at the top; do not tweak rcParams per script.
• Colorblind + grayscale safe: use the palette + redundant marker/linestyle so every series survives a B&W print and deuteranopia.
• Units + N + uncertainty on the figure: axis labels carry units; N annotated; CI band / error bars shown (§0b axis 3).
• Acceptance test (run it, don't eyeball it): before declaring done, run the gate python scripts/critique.py scripts/figN_<name>.py — it enforces the §F checklist + §G anti-patterns mechanically (must report no FAIL) and prints the four judgment prompts. Then answer the four §0b axes by hand using references/figure-critique.md (the mechanical floor is defense; the four judgment ticks are the offense). A figure that merely "plots correctly" — or merely passes the script — is not done.

Optional front-ends / references

• §M Origin — allowed only behind the M1–M6 guardrails (same-source CSV, journal .otp template, vector three-format, reproducible matplotlib fallback for key figures). Hero figures never go through Origin.
• §L external template library — if <TEMPLATE_LIB> is set in CLAUDE.md, use it as a design reference (imitate composition/palette/chart-type, redraw original); never paste foreign output into the paper.

Captions

After the figure passes the critique gate (references/figure-critique.md), write the caption with references/caption-and-quality.md — the caption states the conclusion, not "Figure shows X"; reference figures in text as "as in Fig. 3", never "the figure below".

Output locations

• scripts → scripts/figN_<short-name>.py (or .tex for TikZ)
• figures → outputs/figures/figN_<short-name>.{pdf,png,svg}
• register each in outputs/figures/figure_manifest.csv (id, path, claim, source data, generation script).