We split them. The agent writing the code never writes the checks it's judged by, and those separate checks caught way more than its own tests did.