A library of the AI safety record: the papers, system cards, safety frameworks and government guidance, searchable to the page. Every citation is checked against the page it names.
Two quotations of the same page
Shevlane et al. say model evaluation is “not the only tool available for AI risk assessment.”
On p. 6, word for word
A summary that drops one word: “Model evaluation is the only tool available for AI risk assessment.”
Not on p. 6The page says: “Model evaluation is not the only tool available for AI risk assessment”
The check is code, not a model: the quoted words must be on the cited page, in order, with nothing skipped. How it works
Try"alignment faking"sabotage evaluationsparse autoencoder*sycophancy NEAR/10 evaluation
- Documents
- 13,902
- PDF pages
- 165,940
- Paragraphs of web documents
- 489,391
- Read in full here
- 3,505
Follow one evaluation across the labs' system cards
Watch an agent research
Does the source it cites support this? “Another important tool is mechanistic analysis, i.e. studying the model’s weights and activations for understanding how it functions (Nanda et al., 2023; Olah et al., 2020).”
8 verified0 page only1 rejected
An early run on a six-document test corpus
CitationBench, the benchmark
Does an agent cite the right page, and quote it accurately?
6 early runs; every number links to its pages
Check a paragraph
Paste text with citations. Each one comes back Verified, Page only or Rejected.
No account
Connect your agent
An MCP server with four tools: search, fetch, verify_quote, check_answer.
claude mcp add --transport http safetypage https://safetypage-api.noahairmet.com/mcpThe shelves
How it is made
The text of record is each PDF's own text layer or the web page's own words; no model transcribes a document. A quote check is code, the same code for the reader, the MCP server and the benchmark. Three quarters of the library is not openly licensed: those documents are searched, cited and quote-checked here and read at their source. How it works · Source on GitHub · The design and its decisions