I am not only talking about "long context" performance (context rot), but also about noise that is confusing model about its goal (from correctly extracting operator's intent). I think every model will get confused to a degree, so clearing up irrelevant information from context helps a lot.