The system generates multiple candidate rollouts with the same model, then conducts separate investigations into claims where rollouts disagree and claims on which they all agree. The investigations use environmental evidence and feed an adjudication step that delivers a revised artifact with a verification record.
The project page reports evaluation across five benchmarks covering documents, spreadsheets, code, and multi-file workspace tasks. With ten rollouts per task, the reported average native benchmark score for Gemini 3.5 Flash rises from 47.2 for a single rollout to 53.4 after evidence-backed revision. The authors report the highest primary selection score in all ten model–benchmark settings. The results describe a specific evaluation using two frontier models; the page also notes added tool calls and latency. They do not establish that the harness improves every task or model.