# Research figures [Computed package demonstrations](computed/README.md) add spatial fields, stochastic ensembles, replicated comparisons and incomplete-data reconstruction, with full numerical inputs and a generated table. They have a separate pinned builder and numerical verification workflow. The website uses SVG exports. Ten architecture/method schematics have editable TikZ sources in [`tikz/`](tikz/); two numerical illustrations use [`gen_research_figures.py`](gen_research_figures.py). Every figure also has a vector PDF for papers/slides and a PNG at its original URL for compatibility. The [figure gallery](https://jcmacdonald.dev/figures/) links the exports and sources. ## Rebuild and check Install TeX Live with `standalone`, TikZ/PGF and Latin Modern, plus Poppler's `pdftocairo`. On Debian/Ubuntu these are supplied by `texlive-latex-extra`, `texlive-pictures`, `lmodern` and `poppler-utils`. Python dependencies for the two plots are pinned in the builder's uv script metadata. ```sh uv run assets/img/research/build_figures.py python3 assets/img/research/build_figures.py --check ``` The schematics can be rebuilt without Python plotting packages: ```sh python3 assets/img/research/build_figures.py --schematics python3 assets/img/research/build_figures.py op_system_slide ``` The second command selects one figure. `--check` verifies source fingerprints, accessible SVG descriptions, static references and the three export formats. TeX builds reject missing glyphs and overfull boxes. Review generated figures visually too: these checks cannot establish scientific correctness or detect every collision between diagram elements. After regeneration, run the repository's full Prettier command and review the SVGs at desktop and mobile widths. Normal Jekyll builds serve committed assets and require neither TeX nor client-side TikZ rendering. Exported glyph paths keep typography independent of browser fonts. SVG titles/descriptions and page alt text provide the textual alternative; the source and PDF retain editable or extractable text. Choose the layout from the relationships being explained. Compilation can branch into several outputs, solver capabilities can be compared in a table, provider interfaces can surround their consumer, and adaptive studies can have feedback paths. Use numbered steps when the figure describes an actual ordered procedure. Share typography and colors without imposing a fixed number of stages or the same diagram structure on different packages. ## Scientific scope `figures.json` records each figure's title, description and source. Keep its claims aligned with the linked package capability guides and page text: - The OP System/Engine schematics explicitly cover PDE/structured-field, nonlocal/integral, compartmental and stochastic reaction models. Shaped arrays and RHS callbacks are numerical representations; the compiler publishes contracts and numerical consumers own discretization and integration. Individual ODE or reaction examples illustrate a particular path. - The package architecture diagrams separate mathematical cores from campaign providers. Backend qualification is method-specific and CPU/GPU evidence is stated explicitly. Cross-axis constraints are metadata in op_system. - `beyond_onehealth` distinguishes public simulation and evaluation tools from op_infer, working inference and calibration research software that remains private, pre-alpha and unreleased. `future_directions_general` maps research aims, including structural evaluation; neither figure reports results from ongoing studies. - `pp_eigentest_schematic` describes the research question and scope of private pre-release work. `vbpca_py_schematic` describes fitting and predictive assessment, without implying universal calibration. - `op_infer_slide` summarizes working fitting, calibration and study-comparison capabilities in a compact graphic for the project card and detail page. Its status remains private, pre-alpha and unreleased; the three capability panels do not specify an inference algorithm or report study results. - `optimal_design` uses synthetic candidate costs `[1, 2.1, 3, 3.7, 5.4]` and utilities `[0.28, 0.45, 0.43, 0.72, 0.80]`. These are illustrative inputs. Design C is dominated; D maximizes utility under the declared cost limit of 4. - `structural_fidelity_comparison` constructs two sigmoid trajectories against an upper prevalence bound over the plotted horizon. A curve remaining below the bound on this interval does not establish how it was constructed or how it behaves elsewhere. The figure computes no predictive scores or uncertainty intervals. ## Choice of tools across the software repositories | Figure type | Source and export policy | | ---------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- | | Architecture, contracts, mathematical methods | TikZ source, SVG for documentation, PDF for publication; small editable SVGs can remain SVG sources. | | Numerical results, diagnostics and uncertainty plots | Keep arrays and analysis in the example's plotting code; export vector SVG/PDF where useful and retain data/provenance. | | Generated model graphs | Use the model/specification as the source; Graphviz or Mermaid can be appropriate for automatic topology. | The review found an editable chemostat operator-split SVG and a numerical comparison PNG in op_engine. trade-study's calibration, Pareto, sensitivity and study plots are generated by its example programs; FlepiMoP2's scenario plots likewise have numerical/R/Python sources. Converting those plots to hand-entered TikZ coordinates would sever their connection to computed results. The compiler documentation primarily uses executable model examples rather than image assets. Reference documentation: [PGF/TikZ](https://ctan.org/pkg/pgf), [Poppler utilities](https://poppler.freedesktop.org/), and [Matplotlib output](https://matplotlib.org/stable/api/_as_gen/matplotlib.figure.Figure.savefig.html).