Figure Studio

Figure Studio turns recorded data and editable recipes into publication assets. Open Figures → Figure Studio. The gallery preserves designs and datasets, while Source, Design and Publish organize each item's workflow.

Choose New figure or New table. Import a workspace CSV/TSV/JSON file, drop a file, paste spreadsheet cells or type a small dataset. Inspect inferred column types, units, missing values and distinct counts before choosing a chart.

Original inputs remain intact. Editing values in a linked dataset creates an editable copy rather than silently rewriting the original file. Studio records source fingerprints, parsing rules, column mappings, preparation steps, theme version and generated outputs.

Use tidy repeated rows when you have seeds or folds:

method,seed,step,accuracy
Baseline,1,100,0.72
Baseline,2,100,0.74
Ours,1,100,0.78
Ours,2,100,0.80

Keeping repetitions lets the recipe summarize them explicitly. Do not discard repeated measurements and invent error bars from a final mean.

Preparation steps include filtering, sorting, derived columns, normalization, group summaries, spreading/gathering columns, smoothing, renaming and joins. Inspect the prepared rows after every analytical transformation.

Derived expressions can describe quantities such as (ours - base) / base or log10(params). Group summaries include mean, median, standard deviation, standard error, count, first and last. Training curves can use an exponential moving average or moving average.

Choose grouping keys deliberately. Averaging across datasets can answer a different question than averaging repeated runs within each dataset. Record the analysis choice in the caption or method when needed. Smoothing is a presentation/analysis operation that should be disclosed rather than mistaken for raw measurements.

Available forms include line, scatter, grouped/stacked/horizontal bars, heatmap, box, violin, histogram and density. Suggestions can provide training curves, final-result bars, methods-by-dataset tables, confusion matrices or scaling plots.

Map columns, then check axis labels and units, series order, scales and ranges. Wide tables can plot several Y columns directly. Small multiples share axes/legend where configured; panels can combine different charts.

For repeated rows, choose the summary and uncertainty: mean with standard deviation, standard error or 95% interval, or median/range where supported. Supplied upper/lower bounds and plus/minus columns also work. Inspect whether error bars or a band communicate the intended quantity.

Click supported preview titles, axis labels and legend entries to edit. Set per-series labels, colors, markers, line styles and visibility. Legends can sit above, below, beside or inside a corner. Log axes require appropriate positive data; fixed ranges should not hide scientifically important observations.

Paper theme controls the palette, typography, dimensions, line/marker weights, grids, backgrounds and stable category colors. Choose single- or double-column dimensions appropriate to the manuscript. Figure-specific overrides remain local to that design.

A theme update makes dependent figures stale so they can be regenerated consistently. Check grayscale/color-vision previews and printed label size. Enlarge text can help when labels would be too small at the estimated paper column width.

The preview is a useful design aid, but inspect the actual PDF at publication size. Typography can vary by export format; vector PDF uses the renderer's supported standard fonts.

Studio tables offer booktabs, ruled or minimal styles, selected/reordered columns, numeric formatting, value ± error, spanning headers and grouping rules. Best/second-best emphasis can operate by row or column with explicit higher/lower-is-better choices.

Do not use automatic highlighting without checking the metric direction. A lower error rate and a higher accuracy score need different rules. For direct spreadsheet-style cell editing and formulas, use the separate table editor.

Publish/render the current recipe to produce SVG, vector PDF and PDF-rasterized PNG as selected. Figure outputs use stable paths under figures/generated/; table recipes produce standalone LaTeX under tables/generated/.

Insert with the available figure, full-width, wrap or input wrapper. A changed design must be published again before its new output is inserted. Booktabs insertion can add the needed package.

A manuscript build performs remote refresh, a consistent data snapshot, incremental asset generation and LaTeX compilation. Required generation failures preserve previous assets and stop the build. Build with last assets is an explicit stale fallback, not a fresh render. Externally modified generated files require a deliberate replace decision.

Duplicate designs, group them by experiment/section, save reusable presets and copy a complete agent brief. Freeze assets when preserving a submission state; frozen items stay pinned.

The studio's agent composer can request a dataset or design change while leaving the studio visible. The assistant should inspect data, preserve the selected paper theme, preview the recipe, resolve warnings, then save and render. It should ask about analytical ambiguities such as which runs belong together rather than fabricate an answer.

Recipes use a deterministic local declarative renderer. They do not execute arbitrary imported plotting code. Keep raw data, transformation choices and provenance available so another author can reproduce the output.