Contribute

You can help in three ways: point out an error, add a work to the corpus — by transcribing it from its scan or translating one we’ve already transcribed — or review a text that’s already here.

Report an error

Every text has a “Report an error” link. Click it to open a pre-filled issue naming the work and passage.

Add a work: transcribe or translate

The two are the same story with one word changed. Here it is in three steps.

1. Pick what to work on, and claim it.

• Transcribing — pick an unclaimed public-domain work (check the claims list and search the catalog first — or browse the journals for original runs worth reviving). It must pass the copyright gate: both the life+70 and the US 95-year rules, with sourced dates.
• Translating — pick a work already in the catalog and a target language. There’s no new copyright check to clear — the original already passed the gate — as long as you translate from our transcription, not from a modern, in-copyright translation.

Then open a quick claim so nobody doubles your effort. Claims are short-lived: if you haven’t opened a pull request within 24 hours the claim expires automatically, so the list always reflects who’s actually working right now.

2. Produce the LaTeX, in whichever way suits you.

• Your own AI chat, no coding — copy our pinned prompt (transcribe-chat.md or translate-chat.md) into your AI app (Claude, ChatGPT, etc.). For a transcription, upload a few scan pages at a time; for a translation, paste our original.tex for the section. The prompt makes it produce LaTeX in our house style.
• A guided skill — run /transcribe <work> <pages> or /translate <work> <lang>. The skill reads the house-style rules, produces the LaTeX, validates it, and opens a pull request for you. See using the skills elsewhere below — it isn’t only for Claude Code.
• Full pipeline, your own API key — technical contributors run the pipeline/ scripts with their own API key (Batch API), which emits full provenance automatically, then open a pull request.

3. Submit — through whichever door fits you.

• Issue form (no coding needed). Send the LaTeX through the “Chat transcription” or “Chat translation” issue form, which records the work, pages or language, and which model/app you used. A maintainer then assembles it into the corpus format and credits you.
• Pull request (open to anyone, not just maintainers). If you’re comfortable with git, you can add the files under corpus/<work-id>/ yourself, run python pipeline/validate.py and the tests, and open a PR with a DCO sign-off. This gives you full provenance and CI checks up front. The skill and the full pipeline take this door for you automatically.

Using the skills elsewhere

The /transcribe and /translate skills are just folders of Markdown instructions under .claude/skills/ — nothing about them is locked to one app. A skill does two things: produce house-style LaTeX, and assemble it into the corpus and open a PR.

• In Claude Code, or any agent with the repo checked out (for example via the Claude Agent SDK), both halves run: you get the LaTeX and an opened pull request, because the tool can read the repo, run validate.py, and use git.
• On a chat surface without the repo (a plain AI chat app, or another platform that supports Claude Skills but has no checkout), only the first half applies. There, use the matching pinned prompt above — it’s the same house-style rules without the automation — and submit the result through the issue form.

Quality status: AI draft → skimmed → verified

Your text starts as an AI draft (machine output, not yet human-checked) and is credited to you. From there it can move up as people read it: skimmed once someone has given it a quick pass, and verified after a full, rigorous check against the source. Each step records who reviewed it, so the quality status always reflects the care a text has actually received.

You can do those review steps yourself — no need to be the original author. Open a work, read it side by side with the linked scan, and:

• Spot something wrong? Use the “Report an error” link to flag the passage — that alone helps the next reviewer.
• Want to promote it? (no coding needed) Fill in the “Review a work” issue form: say which text you read, the level you reached (skimmed or verified), and what you checked. A maintainer records you in the reviewers list and advances the status.
• Prefer a pull request? Edit the work’s provenance.yaml directly: add yourself to reviewers with the level and date, a short note on what you checked, and raise the artifact’s status to match. A maintainer confirms it against the scan and merges.

Even a partial check is worth recording — just say in your note what you did and didn’t cover.

Values

• Notation is faithful, presentation follows house style. We keep the author’s own symbols, wording, and choices exactly as printed. How the same math is set (spacing, display vs inline, labels, headings) follows our house style so the corpus reads consistently. See HOUSESTYLE.md for more details.
• Public domain only. New works must pass the copyright gate (both the life+70 and US 95-year rules) with sourced dates.
• Credit & licensing. Contributions are released under CC0 via a sign-off; you are credited in the work’s provenance and on its page.