It would depend on how things are architected, but you could definitely test the components of your example in isolation (e.g. spawn test, get the movement vector in response to an enemy within a certain proximity, test that the state is set to attacking, whatever that looks like). I don't disagree that it's a hard problem. I run into similar issues with systems that use ML as some part of their core, and I've never come up with a satisfying solution. My strategy these days is to test the things that it makes sense to test, and accept that for some things (especially dynamic behavior of the system) you just have to use it and test it that way.