there's absolutely no reason to think Claude produces absolute best token distributions or that slight adjustments would be noticeable.

In fact we know it's not that good because we can often tell Claude's writing apart from human writing.