ccfa-paper-figures

name

ccfa-paper-figures

description

Create, redesign, review, or export publication-ready data/statistical figures and schematic-led architecture composites for CCF-A and top CS venues, especially AI, cybersecurity, software engineering, PL, systems, architecture, and networking. Use for ACM/IEEE/USENIX/NeurIPS/ICML/CVPR/ICCV/ACL/AAAI/IJCAI/ICSE/FSE/ASE/ISSTA/PLDI/POPL/OOPSLA/SOSP/OSDI/CCS/NDSS/S&P/SIGCOMM/NSDI-ready result plots, ablations, benchmark summaries, CDF/CCDF, latency-throughput, scalability, distributions, heatmaps, confusion matrices, scatter/Pareto, stacked-area, radar, qualitative grids, model/data-flow overviews, system architecture composites, security threat-boundary workflows, SE pipelines, Matplotlib/Seaborn/PGFPlots workflows, and camera-ready audits for vector export, embedded fonts, grayscale/colorblind safety, and final-size legibility.

CCF-A Paper Figures

Produce figures as paper artifacts, not slide art. Prefer reproducible source plus vector-first exports. Treat every figure as a claim-supporting object: it should answer one research question, survive one-column scaling, and remain interpretable in grayscale.

Current scope covers data/statistical figures plus Nature-style schematic-led composites: one dominant model/system/workflow/architecture panel, often with subordinate quantitative evidence panels. Use Matplotlib primitives and audited/user-provided content for these schematics. Do not make Graphviz, Mermaid, D2, or diagram-first workflow languages the default path.

Source Discipline

• Use official venue, publisher, or author-kit instructions first.
• Use references/source-audit.md when a rule needs justification or when adding new web-derived material.
• Use references/public-material-audit.md before using any public figure, screenshot, dataset, benchmark table, icon, or architecture material as a reusable material source.
• Use references/extension-gates.md before extending venues, templates, scripts, examples, or bundled assets.
• Do not copy published paper figures into the active skill. Derive rules from observed data-figure patterns and cite source pages instead.
• Do not add self-drawn example figures to the reusable library. Generated figures are task outputs, not reusable examples, unless the user explicitly supplied and approved them.
• Do not invent numeric results, benchmark labels, architecture components, system names, workflow stages, screenshots, or paper examples. Require user-provided content, audited public material, or permission-cleared paper data/material.
• For architecture/schematic figures, draw only the user's described design or audited source material. Public paper figures remain link-only unless redistribution rights are explicit.
• Reuse the local Nature-style figure rules only where they generalize to CS data figures: editable SVG/PDF, restrained palettes, non-redundant panels, hero-panel hierarchy, and final-size readability.

Resource Map

• references/venue-map.md: CCF-A venue families, publisher profiles, and data-figure defaults for AI, cybersecurity, software engineering, PL, systems, architecture, and networking.
• references/figure-decision-guide.md: choose Matplotlib/Seaborn or PGFPlots for data figures.
• references/domain-patterns.md: domain-specific result-plot patterns and integrity checks.
• references/matplotlib-patterns.md: publication Matplotlib conventions adapted from mature Nature-style resources.
• references/plot-api.md: reusable helper functions, CSV contracts, and export helpers.
• references/chart-cookbook.md: chart-family recipes for line, bar, dot-interval, CDF, heatmap, scatter, area, radar, image-grid, and multi-panel data figures.
• references/schematic-patterns.md: Nature-style schematic-led composites for CS model/data-flow, system architecture, security overview, and SE workflow figures.
• references/quality-checklist.md: final readiness checklist for data figures.
• references/source-basis.md: rule-to-source traceability, separating official policy from operational synthesis.
• references/source-audit.md: audited official sources and material-retention decisions.
• references/public-material-audit.md: strict audit ledger for public materials and reuse status, including audited architecture materials.
• references/extension-gates.md: mandatory gates for extending venues, materials, templates, scripts, gallery assets, reference material, and Stable status.
• references/paper-exemplars.md: curated link-only paper references for data-figure inspiration.
• scripts/figure_audit.py: lightweight export audit for PDF/SVG/PNG/JPEG/TIFF.
• scripts/suggest_showpiece.py: profile a supplied CSV and emit a high-information figure plan plus a self-prompt for vague data-first requests.
• scripts/scaffold_figure.py: create a venue/domain-aware data-figure workspace from existing plotting templates.
• assets/ccfa_matplotlib.mplstyle, assets/templates/ccfa_plot_helpers.py, assets/templates/matplotlib_result_plot.py, assets/templates/matplotlib_schematic_composite.py, assets/templates/tikz_pgfplots.tex, assets/gallery/, and assets/chart-atlas/: data-plot and schematic-composite source templates plus migrated Nature baseline previews.
• assets/reference-materials/: audited public architecture/process/security materials. Treat these as reference material, not preapproved reusable templates; check public-material-audit.md before use.

Workflow

1. Parse the user's targeting information before designing:
   • venue: exact conference or journal if named, such as CCS, ICSE, CVPR, NeurIPS, USENIX Security, S&P, PLDI, or OSDI.
   • domain: AI, cybersecurity, software engineering, programming languages, systems, architecture, or networking.
   • stage: submission, rebuttal, camera-ready, journal revision, slide reuse, or unknown.
   • figure type: result comparison, ablation, CDF/CCDF, scalability, latency-throughput, distribution, heatmap, confusion matrix, robustness, calibration, qualitative result grid, or case-study data summary, schematic-led composite, model overview, system architecture, security/threat-boundary overview, or SE workflow/pipeline.
   • style intent: strict/compact, ordinary camera-ready, or showpiece/high-impact/most visually expressive. If the user gives only data and a vague style request, infer showpiece only when the data are rich enough to support a non-redundant multi-panel figure.
2. Choose the strictest profile from exact venue first, then domain, then publisher family. Load references/venue-map.md for venue mapping and references/domain-patterns.md for domain patterns. If unknown, use the strict profile: PDF/SVG source, 3.5 in single-column or 7.16 in double-column, common embedded fonts, colorblind-safe palette, and grayscale readability.
3. For vague data-first requests, profile the CSV before choosing the chart:
   • Run scripts/suggest_showpiece.py <csv> --venue <venue> --domain <domain> --style showpiece when the user says things like "make it high-quality", "more impressive", "Nature-like", "show the data well", or gives only a CSV plus this skill.
   • Use the emitted plan as an internal self-prompt. Do not ask the user to choose a chart unless the data are too small, ambiguous, or missing the claim-critical variables.
   • Prefer the richest non-redundant data page that the CSV supports: matrix-led landscape, benchmark landscape, Pareto/trade-off composite, distribution-led composite, or scaling composite.
   • Reject forced complexity. If the CSV has too few rows or only one meaningful variable, choose a clean single-family chart and explain why a showpiece layout would be misleading.
4. Classify the figure:
   • Tabular experimental data: use Matplotlib/Seaborn with assets/ccfa_matplotlib.mplstyle. The bundled Matplotlib template supports line, bar, dot-interval, cdf, heatmap, scatter, area, radar, and image-grid CSV contracts.
   • Schematic-led architecture/model/workflow composite: read references/schematic-patterns.md and use assets/templates/matplotlib_schematic_composite.py. Put the schematic in a dominant hero panel and place optional quantitative evidence below it. The JSON spec must come from user-provided, audited public, or permission-cleared content. Use semantic kind slots for components, arrows, and boundaries when the architecture has data, control, attack, defense, runtime, storage, human, external, deployment, or trust roles.
   • LaTeX-native numeric plots: use PGFPlots when paper font consistency matters most.
   • Dense heatmaps, qualitative result grids, UI/image examples, or screenshots: keep raster content at submission DPI, add labels in vector overlay where possible, and avoid invented or cherry-picked examples.
5. Create a figure folder when generating assets: figures/<figure-id>/source, figures/<figure-id>/exports, and figures/<figure-id>/data when data are available.
6. Export at least one editable source and one publication file:
   • Preferred: .pdf plus .svg for vector figures.
   • IEEE-compatible: .pdf, .eps, .png, or .tiff; avoid JPEG for plots.
   • ACM TAPS-compatible: .svg, .pdf, .eps, .png, .jpg, .tiff, or .emf.
7. Audit final exports with scripts/figure_audit.py where possible, then inspect visually at final paper size. Do not call a figure ready until small text, legends, color encodings, uncertainty notation, panel labels, and all text placement are readable with no overlaps. On Windows, write audit logs with UTF-8, for example pipe output to Set-Content -Encoding UTF8; avoid UTF-16 redirected logs that become hard to read on GitHub.

Data-Figure Rules

• Use vector graphics for line art, charts, plots, axes, labels, and legends.
• Use color plus redundant encoding: marker shape, dash style, hatch, annotation, or direct label.
• Avoid jet, rainbow heatmaps, 3D chart effects, decorative shadows, low-contrast text, and screenshot-only evidence.
• Do not accept overlapping text. Axis labels, tick labels, legends, annotations, panel letters, callouts, direct labels, colorbar labels, and in-cell heatmap text must not collide with each other or with dense marks at final paper size.
• In tight multi-panel figures, fold panel letters into left-aligned panel titles when standalone panel-letter text would collide with titles, axes, or neighboring panels.
• Use direct labels only when they are spatially stable and sparse. For ECDF/scatter panels with nearby curves or points, switch to a compact outside/in-corner legend instead of forcing labels onto the data.
• Treat colorbars as text-collision risks: give tick labels and unit labels enough padding, or move the unit into the panel title/caption when the colorbar label touches ticks.
• Never let the smallest visible text fall below 6 pt/px in SVG exports; prefer 7-10 pt and use 6-6.5 pt only for dense but still readable tick labels.
• Keep final-size axis/tick/legend text around 7-10 pt unless the venue template requires otherwise.
• Put units in axis labels, define acronyms in the caption, and keep captions out of the image.
• Make every multi-panel figure non-redundant: main result, mechanism, boundary, robustness, or case panels should each answer a different question.
• For ACM submissions, prepare figure descriptions separately from captions when applicable.
• For cybersecurity data figures, define attack success, detection rate, FPR/FNR, overhead, and assumptions in labels or captions.
• For software engineering data figures, show sample size, benchmark identity, project splits, and statistical uncertainty when relevant.
• For AI data figures, show train/validation/test separation, benchmark versions, seeds or variance, and avoid qualitative cherry-picking.

Schematic And Architecture Rules

• Use a dominant schematic/overview panel when the central claim is mechanism, architecture, workflow, or data/control flow; allocate about half the figure height to it when supporting plots are present.
• Draw semantic elements only: components, stages, data stores, trust boundaries, tensor/data/control flows, attack/defense roles, or evidence callouts. Avoid decorative icons and generic clip art.
• Reuse the schematic palette in subordinate plots so the page reads as one claim, not as a dashboard.
• Keep source editable: prefer Matplotlib patches, text, arrows, and raster image plates only when the underlying image is user-provided or permission-cleared.
• Do not copy architecture diagrams from papers into generated outputs unless the user has reuse rights. If using public architecture material as inspiration, cite it and keep it link-only unless public-material-audit.md says bundling is allowed.

Tool Preference

• Matplotlib/Seaborn: empirical plots, ablations, distributions, error bars, Pareto/frontier charts, CDF/CCDF, latency-throughput, robustness, calibration, and attack-success curves.
• PGFPlots/TikZ: LaTeX-native numeric plots and compact math-heavy result plots. Keep CSV data external and do not invent values.
• Raster plus vector overlay: qualitative result grids, UI screenshots, dense image evidence, and camera/image panels when the underlying material is permission-cleared.
• Matplotlib schematic composite: model/data-flow overview, system architecture, security workflow, threat-boundary overview, and software-engineering pipeline figures when the content is supplied or audited.

Do not select Graphviz, Mermaid, D2, or diagram-first workflows as the default. Use them only when the user explicitly requests that source language and accepts that it is outside the Nature-style Matplotlib composite path.

Deliverables

When creating a figure, provide:

• Source file(s): .py, .tex, .csv, .svg, or other reproducible data/source files.
• Export file(s): at least .pdf or .svg; add .png/.tiff when raster submission is needed.
• A short caption draft, and an ACM-style figure description when useful.
• A readiness note: venue profile used, export formats, audit result, data provenance, and any checks that still require the manuscript template.