I agree about anything that tries to assign likelihoods that can be quantitatively operated.
I think it's useful to specify where we expect boundaries to be (i.e. which "escalations" are escalations that are supposed to be prevented and which are totally fine), and what aspirational assumptions we are making (e.g. that we assume that a particular escalation is going to be fixed, and assume it already is when thinking about things we want to be doing that will only bear fruit in a more distant future).
I've found many cases of mostly wasted effort due to lack of knowledge about where boundaries could be, or due to disagreements about where they are. I've also experienced lots of frustration caused by people's refusal to express these in _very precise_ terms.
@jeffw
I think the most important part of that is agreeing on what escalations are fine. It's both very useful, and there's actually a reasonable discussion to be had: usually the escalations have to be fine, because they are composed of pieces exercised in normal operations, strung together. If we want to declare that not to be fine: which of these do we break up and how?