The bugs that look like success: quality checks for AI-generated learning content