AI agent eval suite: can AI write trustworthy test cases?
15 short requirements, two AI models, 3 runs each, and a rubric written before the first run. On points both models passed. With one gate ("fail if the answer states as fact something the requirement doesn't say"), only 5...
