I don't think "non-determinism" is the right metric.

Unpredictability is.

I can have a process that is formally 100% deterministic, but that doesn't help me if I have no clue as to what inputs produce desirable outputs.

And it's not that unpredictability isn't bad enough as is, but with an AI it might produce something great, and it might ... not.

And as long as we don't want to touch the code, that same unpredictability is also in any mitigation measures.

When there is some level of predictability in my mitigation measures, I have some confidence that they will eventually converge.

When there isn't, I don't. Doesn't mean that it can't or won't. Just that my ability to get it to converge is at best limited.