That’s exactly the connection I hoped to make. If the contract stays in the reviewer’s head, the agent and reviewer can end up working from different definitions of “done.”

I’d take it one step further: define the evidence when assigning the task. Which behavior must change, what must stay compatible, and which check would fail if the fix were wrong? That gives the agent a clearer target and the reviewer something concrete to challenge. A second agent’s review can help, but it still needs to point to a reproducible issue or a missing check.