Comment by azertyvode

How do you prevent the agent from reward-hacking the hidden eval? e.g. writing training data that effectively leaks the eval distribution rather than teaching a general skill?