I'm taking the gap > 2 * epsilon condition back to my own evals, since it gives a concrete check in place of average error. I've seen the same pattern with LLM judges scored on overall agreement, where a judge matches humans on most items and still flips the close calls that decide which prompt ships. Your two-action example, where the (10, 11) offset keeps the ranking and (0.6, 0.4) flips it, is the one I'd show anyone reading MSE as a ranking metric.