The non-monotone result is the load-bearing one, because it kills a habit everyone has: reading a ranking at one data budget and assuming it carries to another.

If A beats B at 1M steps and the order flips at 10M, the 1M leaderboard was never a ranking of the algorithms. It was a ranking at 1M, with the data budget as a hidden variable nobody reported.

So the fix isn't a cleaner single number. It's refusing to publish a point where a curve is required.

In your large-scale runs, was the crossover gradual, or did pairs cleanly swap order? Gradual gap-narrowing is a nuisance. A clean inversion means a slice of published SOTA is a budget artifact, not a result.