1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65# Working in this repository
This repository is the public benchmark for AI-generated alt text. See
[README.md](README.md) for what the project is and what it measures.
Accessibility is the number one priority here, and it applies to our own
artifacts, not only to the thing we are measuring. Two hard requirements govern
every change.
## 1. Everything is plain text
Every file added to this repository must be plain text, readable and editable in
any text editor, any terminal, and any screen reader without special software.
- Documents: Markdown (`.md`) or `.txt` only. Never commit Word, PDF, slide
decks, Notion or Google Docs exports, or any other binary or proprietary
document format.
- Data: `.csv`, `.json`, `.jsonl`, or `.yaml`. Never spreadsheets, database
dumps, or binary serializations.
- Results: plain text first. Publish the numbers as text; a visualization is
generated from that text and is never the only representation of a finding.
- Never put information only in an image. No screenshots or diagrams standing
in for content that belongs in text.
- No HTML, no embedded styling, no ASCII art in documents.
- Wrap prose at roughly 80 columns.
- Use semantic headings, nested in order, with no levels skipped.
- Prefer lists to tables. If a table is genuinely clearest, keep it narrow
enough to read unrendered.
- Use only widely supported Markdown. Do not use syntax that degrades into
literal punctuation when rendered, such as definition lists (`term` followed
by `: definition`), which GitHub does not support.
- Write link text that describes the destination. No "click here", no bare URL
used as a description.
- Use emoji and symbols only when they are the subject under discussion, never
as decoration and never as the sole carrier of meaning.
The one exception is the benchmark corpus itself: the images under test, kept in
`projects/corpus-construction/corpus/images/`. They are artifacts, never carriers
of documentation, and each is recorded in plain text with its source URL,
taxonomy category, surrounding context, and gold standard alt text. One tool
writes that directory and nothing else may; see
[its README](projects/corpus-construction/corpus/images/README.md).
## 2. Code is held to the same standard
- Read and write plain text formats.
- Make output usable without a graphical interface. A terminal and a screen
reader are the baseline, not a fallback.
- Store configuration and prompts as plain text so they are reviewable in a
diff.
- If code ships an interface, it must be accessible: semantic HTML, keyboard
operable, screen reader tested, never reliant on color alone. Every image an
interface renders needs correct alt text, judged by the same criteria this
benchmark scores.
- Do not add a dependency whose only purpose is producing a binary artifact
where text would do.
## Before you commit
Check that new and changed files satisfy both requirements above. If a change
seems to need a non-text format, say so and ask rather than committing it.