Scientific Figure Master

The only textual input is the paper's own caption; the data is the raw qPCR Ct workbook from its public deposit. From that one sentence the skill extracted the error bar (SD), the central statistic (mean), the test (Student's t-test at each time point), the star scheme (*<0.05, **<0.01, ***<0.001) and the eight panel letters — then computed ΔΔCt fold change per replicate, ran the tests, and rendered the figure.

The published panel is shown beside the generated one so you can judge the result yourself. All eight gene panels' bar heights and all 32 significance-star tiers agree with the published figure.

Nothing but caption text drove these. A — grouped qPCR bars: the caption named genes, treatments, days, SD, two-way ANOVA with Fisher's LSD and the star scheme; the pipeline found the columns, computed per-day ΔΔCt and placed the brackets. B — volcano: the caption gave the contrast and thresholds (|log2FC| > 1, FDR < 0.05), built from a public count matrix.

C is the honest case. On data it had never seen, the caption never named a chart type — "median with interquartile range; Mann-Whitney U test" was enough to infer box plus individual points. The test returned p = 6.7 × 10⁻⁶, and the figure still shows two stars, because that caption defined no higher tier. The skill will not invent a four-star result.

Caption parsing is measured, not asserted. The bundled eval set holds 20 real captions — from the papers above, from eLife 2023;12:e83345 (CC BY 4.0), and from common clinical and omics caption idioms — scored on the fields a competent reader would call unambiguous.

Run it yourself after install with python3 tests/run_caption_eval.py. The set is also a regression guard in the unit suite, so a parser change that breaks a caption fails the build. Captions that genuinely under-specify (a bare figure title) are reported as unspecified fields rather than silently guessed.

Before trusting a figure engine, check its arithmetic. Each panel below was rebuilt from public raw data and compared against the published values. BioFactors 2025;51:e2131 Figure 6E — all eight regression panels reproduce every printed digit: slope, intercept, R² and Wald p (for example Ogt: y = 0.7639x + 0.2117, R² = 73.28%, p = 0.0067).

BioFactors 2021;47:992 Figure 1A — the recomputed viability ANOVA lands exactly on the published p = 0.0003, F = 6.504; Figure 3's split Spearman/Pearson heatmaps reproduce every readable published p-value to four decimals using exact-permutation Spearman, and its PCA reads PC1 41.91% / PC2 30.26%. PeerJ 2020;8:e10244 Figure 7 — the FRAP two-way-ANOVA variance decomposition matches (18.4/68.0/3.6 versus the published 18.52/67.78/3.72%).

On an independent dataset it had never seen, the bioinformatics path recovers the published SARS-CoV-2 Calu-3 interferon signature from raw counts (IFIT2, IFNL1-3, CXCL10, TNF, IL6; PC1 separates the groups at 96.3%). Wiley figures are linked rather than embedded — open them beside the generated panels shown here.

The bundled examples/gallery.py renders nine core archetypes offline in seconds — your install self-check. The catalogue covers 22: dot-on-bar and dot-strip with LSD/FDR brackets, grouped two-factor bars with a two-way-ANOVA interaction p, box and violin, replicated and single-series time courses, right-axis-FDR series, regression with 95% CI, split correlation heatmaps with exact p-values, clustered heatmaps, radar profiles, PCA, volcano with decluttered gene labels, enrichment dot plots, survival, forest, broken axes, blot montage with densitometry, qPCR ΔΔCt, kinetic model fits and Seahorse OCR/ECAR stress tests.

Style is a parameter, not an opinion: the default is a neutral publication look that every journal accepts, while prism (the GraphPad dialect) and nature are opt-in presets chosen from your target venue. Column widths for Nature, Science, Cell, PLOS, IEEE and Elsevier are built in.

No decorative statistics: every star comes from a named, re-run test, and the caption's own legend defines the vocabulary — the skill never invents a tier the caption does not define. No AI-drawn data: numbers are always plotted by code, and image models are reserved for illustrations and schematics.

It contradicts the caption when the data does. If a caption claims a significant increase and the test returns p = 0.2, the output says so instead of drawing a star. Fields a caption leaves unstated are reported as unspecified, filled explicitly and surfaced — not buried. And where a public deposit disagrees with its own paper (post-publication revisions), both values are reported rather than one being tuned to match.

23 files · 338 KiB unpacked · standard Agent Skills layout

Emerged in AROS-core on 2026-08-12 by merging three in-brain figure skills (data_visualization, scientific-illustration-generator, scientific-schematics) and hardening the result over three adversarial rounds: figures rebuilt from public raw data for three peer-reviewed papers, an independent GEO dataset, figure-convention audits of eLife open-access papers, and a 20-caption extraction eval. v1.1.0 refactored it from a reproduction tool into an intent-driven engine — venue specifics became style presets and the reproduction log moved to references/validation_evidence.md as evidence rather than doctrine. Workflow patterns adapted from scipilot-figure-skill (MIT © Haojae); journal constants from scientific-visualization (MIT © K-Dense Inc.).

b5378310f385929f3c47f90be0b82f264550dbf6210a0fcd27f0342308bd7ffc