Practical recommendations include: shift from pairwise win rates to rubric-based or pointwise verdicts; explicitly declare the reference baseline before claiming any gain; make "actually look at what happened" cheap; and do not trust a single aggregate score.