I only spent one day at the recount, but was able to observe more precincts than many people were able to. One thing that I noticed was that when the counts were done properly, there was a small shift to Clinton from the undervotes. I was noticing 2-6 per precinct. If there was consistency, it would add up to almost 7,000 votes in Wayne county alone. Richard Hayes Phillips had been saying that we need to look at the undervotes in Wayne, especially outside of Detroit.
Please understand that this is a theory based on a handful of cases. I can only say that as the day went on, I could not disprove this theory.
The handful of undervotes that I saw were actually x or check marks instead of filled in properly.
Here is a copy of the ballot that is similar to the one in Wayne: arthur1c
Here are the instructions on “How to Mark Your Ballot”, from the Michigan Secretary of State website.
Mandated_Ballot_517304_7We were told to count any mark inside the oval as a vote. A circle or arrow to an oval or candidate was not allowed.
One of the clerks that was assisting us, said that the scanners should be able to read any of these marks. However, what I observed was that the marks were only read correctly for the Trump ballots.
For example, while this is what the instructions showed to do: 
I saw examples like these be counted accurately for Donald Trump:
While these were not counted accurately by the machines for Clinton:
At one point a Trump observer asked about the a Trump ballot with a check and he was surprised to realize it was already counted.
There is a third possibility that I did not see to consider, and that is a squiggly line or smaller mark in the oval.
So my theory is that it is possible that the scanners were adjusted to read one candidate while not reading another with these incomplete marks. Compared to things like Cross Check, it would be an innocuous dirty trick. While it wouldn’t make a huge difference in an election, it could be used as a part of an overall strategy to tip an election.
Again, this is only theorizing, with a handful of samples. More work would need to be done to prove or disprove this theory. Better yet access to ESS’s source code.
Update 1: An observer in Wayne county did observe an undervote go to Trump.
Update 2: Jan Bendor from Michigan Election Reform Alliance sent me this:
In Michigan, none of the 7,000 scanners has an adjustable scanning function. They are all 13-14 years old. This feature has not been available until recently, and it is highly problematic.
We believe there are two explanations for the undervotes so far. A programmed instruction on the machine’s memory card to fail to read the Presidential race in a certain percentage of the precinct’s ballots; and a disturbing sign of tampering with absentee ballots. In Michigan, 80 of 83 counties use outside service companies to program the memory cards, all of them connected to the major scanner companies, so this form of rigging would be very easy to insert. The outside techs are NOT sworn officials.
We are seeing a small number in numerous absentee count board precincts of totally blank ballots, around 2 per precinct. These could have been inserted into the return envelopes, replacing original voted ballots.