Paper Figures & Tables

Use this as the single entrypoint for paper artifacts: tables, precise data figures, generated conceptual figures, captions, and artifact QA. It combines public integrity defaults, adaptable author preferences, data-visualization judgment, conceptual-figure planning, and experiment artifact checks.

Use paper-policy for manuscript integrity and applicable project/venue requirements. Select presentation guidance from the artifact’s needs and adaptable author preferences.

For authorization and internal handoffs, follow ../paper-policy/references/authority-model.md (Task authorization and internal handoffs). Continue authorized local work across skill boundaries without another user invocation; preserve audit-only and explicit approval-before-editing scopes.

Managed Research Workspace Preflight

Before creating or changing persistent paper artifacts, resolve a valid PROJECT-WORKSPACE.md for the selected project, or the .research-workspace.yml link when operating in its independent paper Git repository. In a managed workspace, put final lightweight exports only in the declared paper repository and update its figure-manifest.yml with the experiment ID, run or data snapshot, generator, source path, and code revision. Keep raw data, logs, checkpoints, and plotting scratch outside the paper repository.

If no contract resolves, use the normal artifact workflow. Do not initialize a workspace merely because an artifact is academic. Route repository setup, asset-provenance repair, or workspace reorganization to research-workspace-governance.

Priority Model

Always load references/priority-model.md first. Load only the task-specific references needed for the current artifact.

Authority order:

1. Core integrity constraints from the sibling paper-policy skill.
2. User-provided data, manuscript facts, approved evidence, and explicit instructions.
3. Verified venue requirements.
4. Applicable hard rules from the shared academic workflow.
5. Writing handoff specs within the authorized task; a generated spec cannot override hard rules.
6. Editorial guidance, adaptable author preferences, and precise-plot advice.

Do not let generic artifact advice override user facts, enabled policy rules, or route boundaries. Presentation preferences do not establish academic correctness.

Task Routing

Task	Load
Decide table vs figure	artifact-routing.md, captions.md
Related Work comparison table	tables.md, captions.md; add dense-empirical-tables.md for a load-bearing multi-axis matrix and layered-capability-matrix.md when 8+ dimensions form explicit semantic layers
Result table, findings index, sample ledger, taxonomy table, risk matrix	tables.md, captions.md, quality-checks.md; add dense-empirical-tables.md when hierarchy or density is high
Precise experiment/data plot from supplied values or artifacts	data-figures.md, captions.md, quality-checks.md; add profiling, chart-selection or recipes only when the task needs those decisions
Journal/venue-specific plot sizing or export	journal-specs.md, visual-qa.md, publication-checklist.md, then the relevant artifact reference
Conceptual figure, Figure 1, architecture, pipeline, threat model	conceptual-figures.md, conceptual-style-reference.md, conceptual-vector-rebuild.md, captions.md, quality-checks.md; add figure-contract.md when a structured spec helps
Statistical summary for artifact creation	source-data-and-statistics.md, then tables.md or data-figures.md
Multi-run result table or plot	source-data-and-statistics.md for run validity, comparability and reproducible aggregation, then the relevant artifact reference
Caption-only task	captions.md plus the relevant artifact reference
Artifact QA or cleanup	quality-checks.md, then the relevant artifact reference

For final figures/tables, source-data artifacts, Related Work tables, or any artifact feeding submission readiness, also load policy-integration.md.

Artifact Constraints And Defaults

• Use source data or user-provided values. Do not invent numbers, baselines, methods, p-values, error bars, or visual trends.
• Every artifact has a clear scientific purpose; related comparisons may share an organized multi-panel figure.
• Every data-driven artifact must identify its source file or state that source data is missing.
• Use booktabs as a clean option when compatible with the template. Natural-width tables are valid; verify final readability under TABLE.FINAL_READABLE.
• Prefer compact tables after pruning dimensions that add little to the claim. Use wider layouts when needed for a fair, readable comparison.
• Create an axis-based comparison table when useful to the argument or explicitly requested.
• Precise data figures must be derived from source data with reproducible plotting tools. Do not generate numeric trends or error bars with an image model.
• Experimental figures require at least 2 meaningful plot panels per row at single-column placement and at least 4 per row when spanning both columns. Follow data-figures.md for panel counting, final-size readability, and conflicts; never fabricate or duplicate evidence to fill the grid.
• Single-column experimental tables have no added size/density quota. Cross-column experimental tables require an information-density review of supported comparison content, occupied width/height, grouping and readable alternatives; see tables.md.
• For new conceptual figures, default to manuscript/placement definition -> ImageGen visual design -> editable SVG reconstruction -> semantic and visual QA. Classify figure type separately from single-column or two-column-spanning placement; require high effective information content for cross-column space. Follow conceptual-figures.md for scoped exceptions, including existing vector edits. Data plots and tables retain their native workflows.
• Keep mathematical symbols, components, arrows, and boundaries faithful to the manuscript. Inspect the actual output; prompt wording is not evidence of correctness.
• Follow the active tools' editing constraints. Preserve source and transformation records for generated or edited artifacts; tool choice alone does not pass or fail scientific quality.
• Prefer captions for interpretation and avoid redundant in-figure titles. Panel labels or headings are valid when they improve navigation.
• Derive source dimensions and text size from the actual placement. Final-width visual QA is mandatory for final figures; source-point sizes alone do not establish readability.
• Prefer vector exports (.pdf and/or .svg) for precise plots and LaTeX inclusion. PNG is acceptable for review packets and generated conceptual figures.
• Keep each data figure traceable to a reproducible script, notebook or shared-pipeline entrypoint; reuse appropriate plotting infrastructure.
• Prefer tables for exact comparisons and figures for trend shape, flow, geometry, and distribution, adapting this heuristic to the artifact and task.
• Writing prose belongs to paper-writing; this skill may polish captions and short artifact callouts but should not draft full paper sections.

Output Contracts

For LaTeX tables:

1. Return package or macro requirements when needed.
2. Return compile-ready LaTeX.
3. Include \caption{} and \label{}.
4. Make metric directions clear and use consistent, correctly rendered notation.
5. Define markers such as \cmark, \pmark, and \xmark only when needed for readability, preferably in a short caption phrase rather than a separate note.
6. Choose compact, dense or semantically grouped layouts as needed, then inspect at actual placement width. For cross-column experimental tables, record the density review from tables.md; single-column small tables have no added density quota.

For data figures:

1. Reuse source and encoding notes; write a compact spec when the figure's complexity warrants it. Formal compliance records are needed only for that assessment.
2. Profile unfamiliar raw data when it helps resolve types, grouping or chart selection.
3. Recommend the chart type from the paper claim and data shape; actively warn when the requested chart hides distribution, uncertainty, or sample size. Record the intended column span and panel count for each row; experimental figures must meet the single-column 2 / cross-column 4 minimums and remain readable.
4. Save or describe outputs: figure.pdf, optional figure.svg, optional figure.png, script, source data path, and caption.
5. Validate source fonts, inspect the actual export at final paper width, and record the actual evaluator. Agent visual inspection is not human sign-off.

For conceptual figures:

1. Establish manuscript-supported components, connections, exact labels, boundaries, and the reader takeaway before visual generation.
2. Record figure type, actual target width, height budget, and an information/layout plan. A wide screenshot does not establish two-column publication placement.
3. Use the available imagegen skill for the design draft, then reconstruct its composition, meaningful icons, color roles, and hierarchy in editable SVG. Treat this as visual design followed by explicit vector reconstruction, not raster auto-tracing.
4. Deliver the SVG source, paper-ready PDF when supported, and a rendered preview; retain the selected design draft, prompt, source/connection notes, and material deviations in the project working area.
5. Independently check scientific correctness against the manuscript and visual fidelity against the design draft, then inspect the export at final placement size. Report incomplete reconstruction or unverified checks accurately.
6. Keep generation provenance through reconstruction. Add generated_conceptual_figure when the final artifact contains image-model output; for a fully reconstructed SVG, record the generated design reference and assess applicable venue rules without implying generation was unused.
7. Reuse existing editable sources for small edits; explicit user/venue requirements take precedence. See the conceptual workflow for tool-unavailable handling and scoped alternatives.

Reference Map

• references/priority-model.md: source priority and conflict resolution.
• references/policy-integration.md: context resolution, artifact evidence, adaptive source fonts, final-width QA, and compliance handoff.
• references/artifact-routing.md: table vs figure selection.
• references/tables.md: LaTeX table style and resizebox rule.
• references/dense-empirical-tables.md: optional design guide for grouped headers, compact empirical matrices, findings-at-a-glance tables, and sample ledgers.
• references/layered-capability-matrix.md: design guide for Related Work capability matrices with semantic column layers, conservative family-row evidence, and grounded coverage-delta highlights.
• references/data-figures.md: reproducible data-driven plotting workflow.
• references/data-profiling.md, references/chart-selection.md, references/visual-pitfalls.md: data profiling, chart selection, and visual-risk guidance.
• references/journal-specs.md, references/visual-qa.md, references/publication-checklist.md: venue sizing, final-size rendering, and visual QA.
• references/conceptual-figures.md: figure classification, placement/density planning, and ImageGen-to-SVG workflow.
• references/conceptual-style-reference.md: distilled sample layouts, semantic palettes, and optional user-supplied visual references.
• references/conceptual-vector-rebuild.md: editable reconstruction, icon fidelity, and separate semantic/visual acceptance checks.
• references/captions.md: self-contained captions and artifact callouts.
• references/source-data-and-statistics.md: source data, statistical summaries, and placeholders.
• references/quality-checks.md: artifact QA checklist.
• references/figure-contract.md, references/plot-patterns.md, references/plot-recipes.md, references/data-figure-style-source.md: detailed plotting resources.

Script Map

• scripts/profile_data.py: profile tabular data before chart selection.
• scripts/paperfig_style.py: reusable helper for paper plots and semantic color roles.
• scripts/setup_style.py, scripts/export_figure.py, scripts/layout_tools.py, scripts/visual_qa.py, scripts/check_figure.py: final-size styling, export, layout, preview QA, and file audit for precise data figures.