it seems as though this feature is not refined. Simple prompts that we do manually with a human get different results every time compared to the automated tests. Also, seems like using the test feature for chats does not work?