I was trying to stress test and our sliding window logic seemed to be failing in irregular windows that might not align perfectly with the candidate set size (which is a majority of cases, besides our settings like (20, 10) on candidates of 20/100, etc.).
This should fix it.
Additionally, some non-breaking changes (for our (20, 10) experiments) when we are dealing with smaller windows (you don't need to fix a 300 token limit).
I was trying to stress test and our sliding window logic seemed to be failing in irregular windows that might not align perfectly with the candidate set size (which is a majority of cases, besides our settings like (20, 10) on candidates of 20/100, etc.).
This should fix it.
Additionally, some non-breaking changes (for our (20, 10) experiments) when we are dealing with smaller windows (you don't need to fix a 300 token limit).