Drawn from source

Claim: a figure drawn as reader-ergonomics' draw relations asks is written in Graphviz: its .dot sits beside the SVG and the page that embeds it, folded under the image. Text figures stay only in files read in a terminal or by a model.

• Grade: convention, untested. The reasons are access and drift, not faster reading.
• Evidence: review, strand 4; position, Known; decision 0007.
• Check: reader-ergonomics' typed-figure and figure-source, which its draw relations names; this principle adds the tool, the place of the source, the folded copy and the exempt files.
• Reader styles: none changes it.

What it asks

• Write a diagram of relations in Graphviz's dot language, render it to SVG, and embed the SVG in the page.
• Keep the .dot beside the SVG, under the same name, beside the page that embeds it and named after it; a folder of figures would group files by type. A model, a terminal and a later editor read the source; a person reads the drawing.
• Fold the source under the image, in a <details> block titled "Graphviz source", so a reader who wants to reuse the figure finds it on the page, closed until clicked.
• Keep a text figure only where no drawing can be shown: AGENTS.md, llms.txt, a comment in code, a commit message. The exempt files are listed in checks/config.toml.
• Render with python3 <reader-ergonomics>/checks/figures/render_figure.py FILE.dot, which stamps the SVG with the hash of its source and refreshes the folded copies.

Why

WCAG 2.2 counts ASCII art as non-text content, which needs a text alternative like any image (W3C 2023a). Its glossary names the 95 printable ASCII characters, so box-drawing characters fall under it by extension. Screen readers read such a figure character by character (CUNY 2026).

A model reads a graph better from text than from an image: GPT-4o found an edge in 96.2% of cases from an adjacency list, against 74.6% from a picture (Zhong 2024). Turning a flowchart into Graphviz or Mermaid source before reasoning on it beat end-to-end models (Ye 2025). The .dot file gives a model that text form.

Limits

• No study compares a text figure with a drawn one, for sighted readers or for screen-reader users.
• A drawn figure can drift from its source. The hash stamp catches that; a figure edited by hand cannot be checked.
• The rule was set by the maintainer on 2026-10-29, after reader-ergonomics argued that a text figure reads in a terminal and to a model exactly as it is drawn. The exempt files answer the first point and the .dot source the second.