There are 2 different methods of determining focus. When they don't agree we generally default to "old" as it is generally more accurate.
New - Linear regression (line intersection) of the V curve.
Old - Weighted lowest local average of 3 best points.
The New method generally works quite well but if a "good" fit can't be obtained (seeing, etc) we default back to the Old. Also if the New method is out of tolerance with the Old we default to the Old as it's generally more robust, at least locally.
If you can duplicate these issues it would be helpful to put together some AF Packs and we can take a look at the data and see if we can adjust this some.
Thanks,
Jared