As far as the actual data, Louisiana is the first state that I looked at that has no vote reductions at the end. So, I have this working theory (with some good data behind it!) that the machines just work with the percentages during the "count". They take the number of real votes during the interval and pad with phantom votes to the R & D sides, as needed to make the target percentage. At the end, they assign the ballots where they padded the numbers, But - if they have phantom voters that are both R&D, they will take the even numbers of each to reduce the totals. That keeps the percentage the same for the win, but removes the extra phantom voters so they have as few as possible to have to fill in and possibly get caught with. There were no R phantoms ever required - plenty of R ballots from real people! So, no reductions made here at the end.
