Sure, there are exceptions. When I wrote ‘normally’ to cover them, I was more thinking of things where it is hard to set good metrics for the thing one wants controlled, or cases where avoiding the spirit of the rule would be too likely.

Nevertheless, I think those are exceptions. I think first order effects tend to matter more than second order effects (that’s why they’re called first order) and not doing good first-order things because of potentially bad second-order things is often wrong.