SafetyPage

CitationBench

Does an agent cite the right page, quote it accurately, and make claims the page supports?

6 runs · updated 2026-11-10

These are early runs on a handful of questions over the six-document fixture library. They show that the harness works end to end; they are too small to compare models.

Correct is the headline: the verdict is right, at least one cited page is a gold page, and no citation is rejected by code.

Code columns are deterministic: the cited page exists, the run opened it, and the quoted words are on it. No model is involved.

Judge columns come from a model, and are labelled so. The judge is claude-code:claude-sonnet-5-5#12160b. Their agreement with human labels has not been measured yet, so read this column as provisional.

Rank is a band, not a position: runs whose 95% intervals on Correct overlap share a band. Intervals come from a bootstrap over questions, clustered by document.

Every number opens the run behind it. Every question in a run opens its replay, and every citation in a replay opens the page. How the questions are made, and how to run it