Figure conventions and reproducibility

Visual conventions

• Language comparisons keep R, Python, PHP in that order. Python is labeled statsmodels or SciPy reference to identify the actual implementation.
• R uses blue circles (#0072B2); Python uses orange triangles (#D55E00); PHP uses green squares (#009E73). Color and shape reinforce the labels.
• Conceptual figures use blue for the reference/population curve, green for the conditional/native curve, orange for an approximation or boundary, and gray for observations. Their legends describe the statistical quantity rather than imply that a curve is an implementation comparison.
• Use a white background, common sans-serif type, restrained grids and explicit quantity/unit labels. Titles state what is shown; captions state the data or approximation, transformations and relevant limits.
• Bar charts start at zero. Benchmark facets use their own labeled linear scales and show median bars plus actual timing repetitions. Missing memory is labeled unrecorded, not plotted as measured zero. A measured time of zero is marked 0* because timer resolution can hide a short fit.
• Log scales are explicit and reserved for quantities spanning orders of magnitude, such as integration errors. The 1e-13 display floor in the quadrature-error plot is disclosed; it never changes stored numerical measurements.
• Error-versus-reference figures use a zero reference line and label any unit multiplier. Agreement plots are not residual or model-adequacy diagnostics.

These conventions use ggplot2’s shared themes and explicit export controls. Each figure is saved as selectable-text SVG and a 160-dpi PNG, with fixed dimensions recorded in the figure manifest. Font family is generic sans-serif; the resolved font can differ across platforms.

Generate figures without running fits or benchmarks

Install the documentation-only R packages:

install.packages(c("ggplot2", "jsonlite", "svglite"),
                 repos = "https://cloud.r-project.org")

From the repository root:

Rscript docs/figures/render.R --all
# Or render an individual family/figure:
Rscript docs/figures/render.R ols
Rscript docs/figures/render.R approximation
Rscript docs/figures/render.R sleepstudy
./scripts/render-docs

The renderer reads saved benchmark/reference reports and notebook-exported points in figures/data/. It does not call estimators, regenerate reference fixtures or run timed workers. A changed numerical report requires refreshed plots; unchanged measurements should not be retimed for a style change.

The manifest records each input path/content MD5, image hashes, dimensions, renderer/theme hashes, R version, platform and direct package versions. The R package record also records installed transitive dependency versions. These are the versions used for the saved figures; installing another version is not a claim of byte-identical reproduction. Data/report provenance remains in its original statistical verification records.

Mathematical profile and posterior plots use points exported by the executable walkthroughs, preserving the visible calculations. Re-executing those .qmd pages updates their small JSON point snapshots and figures. Other figures read existing result reports. Jupyter embeds SVGs with namespaced clip IDs so figures on the same page do not interfere with each other.

Ordinary website rendering reuses frozen outputs and saved images. Neither R nor Python is required to browse the committed HTML. Re-executing computational notebooks still requires the documentation environment.

Existing figure coverage

The common renderer covers all five fixed/Gaussian-intercept benchmarks, the logistic-mixed benchmark, v0.1 preservation comparisons, Gaussian likelihood profiles, published-data numerical agreement, sleepstudy random slopes, logistic-mixed integration error, node sensitivity and conditional posterior shape. SVG and PNG downloads are retained alongside their saved input records.