An invented software company is about to send a tender answer with 26 promises, each citing an annex. Ten of them do not hold. Claude checked it three ways in Brinvik's lab: the tender alone, the tender with its annexes, and the annexes plus a cite-check skill. Pick one and press the button.
The same tender and the same request every time; only what Claude was given changes. The bar is the share of the ten errors caught. Click a run to replay it above.
Invented documents: Lindholt Systems, Ravnsø Forsyning and Bredmose Vand do not exist. Every reason shown is Claude's own output from the test run on 29 October 2026 through Brinvik's lab-api workflow. Costs are list prices for the tokens each run used.