A new research system tracks not just whether an AI agent reached the right answer, but whether it had the authority to act and proof its work was actually finished.
Researchers built Matrix, a deterministic causal-state layer that logs which authority approved each agent decision and which facts a given task depended on. It verifies completion evidence before treating a task as done, and when an underlying fact changes, it invalidates only the downstream work tied to that change instead of the whole workflow. In controlled comparisons, agent workflows run through Matrix and those run without it usually reached the same outcomes. The difference was in the paper trail: only the governed version preserved evidence of who approved what, refused to close out unsupported work, and limited any redo to the tasks actually affected.
That distinction matters more than another accuracy score. Agentic AI moving into institutional settings needs someone to answer for who authorized an action and whether "done" actually means done, and that's a compliance problem separate from whether the agent got the right answer.
The system also has a real weak spot. In a role-separated transfer test, a strict completeness rule meant to catch unverified work instead over-blocked synthetic data packets that legitimate processes had produced outside its own authoring context - a reminder that governance layers built to catch fraud can end up punishing normal work too.