Google and the University of Hong Kong release VeriCUA paper: Using machine state to verify task completion claims by computer agents

The VeriCUA paper from Google and the University of Hong Kong argues that task completion reports made by computer agents cannot serve as evidence; instead, saved files and application data should be examined. VeriCUA first explores application build requirement lists, then runs screen or code checks to verify items one by one; on ScienceBoard, its false positive rate is 5.4%, lower than the 24.8% for VLM reviewers and the 21.7% for generated script validators, and it rejects approximately 80% of fake completion reports across three online environments.