AI System Cards

A readable archive of AI model system cards. Companies usually publish a model's system card as a long PDF, and long PDFs are a pain to read. This project converts each one into a faithful, responsive web page — every sentence, table, figure, and footnote, with deep links back to the source PDF — plus clean markdown for machine consumption (card.md per card, llms.txt index).

Site: https://malob.github.io/ai-system-cards/

The conversion is mechanical — no LLM transcribes or edits the content. A PyMuPDF "oracle" plus docling extract ground truth from the PDF; a block compiler assembles faithful markdown; independent invariant gates verify it against the oracle (text, links, styling, structure, tables, figures, footnotes, page markers); and an Astro site renders it to static HTML with PDF deep links, sidenote footnotes, per-page social-preview images, and search.

Status

Built and validated end-to-end on its first document (Claude Fable 5 & Mythos 5, 319 pp.). The site renders any card under cards/, but the conversion pipeline is heavily specialized to that one document — likely not generalizing cleanly even to Anthropic's other system cards, let alone other companies'. It's a strong starting point, not yet a general tool. The next step is to convert a second document and find out whether one shared pipeline can serve many or each will need its own.

Layout

cards/<vendor>/<slug>/
  meta.yaml           # title, slug, vendor, models, release_date, source_url,
                      #   source_pages, description, and the chip-label vocabulary
  source.pdf          # the original PDF
  style-manifest.yaml # per-card visual vocabulary (chip/highlight fills, role colors)
  sections/*.md       # faithful markdown, mechanically generated, ordered by filename
  assets/figures/     # figure images extracted from the PDF (pPPP-K.png)
  extracted/          # per-card extraction data: figure map + image inventory; renders (gitignored)
pipeline/             # the conversion pipeline (extract → assemble → verify)
site/                 # Astro site rendering cards/ to static HTML (deploys to Pages)
docs/                 # design notes, decision log, and re-runnable experiments
CLAUDE.md             # operational playbook for an AI coding agent (see below)

Working on this repo

This project is built and maintained with an AI coding agent. To convert a new system card, or to fix or improve an existing one, open the repo in Claude Code (or your agent of choice): the working instructions — how the pipeline runs, the commands, and the card-generation / "adding a card" workflow — live in CLAUDE.md, with the design rationale and decision log in docs/. Issues and PRs are welcome, including the pipeline generalization itself.

A note on content

The documents reproduced here are published by their respective companies for public consumption; this archive reproduces them faithfully and links each page back to the original PDF. All document content belongs to its publisher.

License

MIT, for the code and pipeline. The reproduced system-card documents under cards/ belong to their respective publishers (see the note above).