About this archive

Scope

The archive covers official model documentation from six labs (Anthropic, OpenAI, Google Gemini, Zhipu GLM, Moonshot Kimi and Alibaba Qwen) for models released between 1 October 2025 and 22 October 2026. It includes system cards, model cards, addenda, safety and risk reports, and technical reports. The Chinese labs and Google's Gemma models usually publish open-weight model cards on Hugging Face or a developer site instead of a system card, so those cards are included as their main documentation. Blog posts are included only when no other official document exists for that model.

Pipeline

  1. Collection. Documents were found on each lab's official index pages (for example anthropic.com/system-cards, OpenAI's Deployment Safety Hub, deepmind.google/models/model-cards, arXiv and Hugging Face) and downloaded in their original form.
  2. Conversion. PDFs were converted to markdown with Mistral OCR. Page markers are kept so that every citation can link back to a page. Documents that exist only as HTML or Hugging Face READMEs were rendered to PDF with headless Chrome. Those PDFs are marked “rendered”.
  3. Extraction. Each document was read against a fixed schema: headline capability benchmarks, a safety framework determination, and 21 safety areas. Every safety area was rated not covered / mentioned / light / substantial and backed by a verbatim quote and a page number. The extraction was done by an LLM (Claude) working directly from the markdown. It can contain errors, so treat the PDF as authoritative.
  4. Analysis. The safety page combines the per-document extractions into trends and gaps. Each claim on it links to the page that supports it.
  5. Replicability. Every named eval was labelled for whether its data is public and whether its judge is public (programmatic, open-weight or a named API model). The replicability page aggregates the labels.

Caveats