What were they thinking?

I've been slowly watching all the ballot images from November's election, at one per second, when I have a chance.  This was a small enough election that I can do this on my own to get a manual count to compare with both TEVS and Hart... I'm through 3800 of our 14000 ballots. 

Everyone needs a hobby, right?

Here's another not-very-easy-to-capture ballot from November's election.  TEVS missed all four votes on this ballot.  I'd bet Hart caught them, but I wouldn't bet my life on it.

Was the voter expressing a lack of enthusiasm, or was it a lack of #2 pencils?

 

A Challenging Ballot

Here is a ballot fragment from Humboldt's Nov 2015 election which proved challenging for TEVS.

Voters are instructed to fully fill-in the rectangles to indicate a vote, and most do so.  This voter used x-marks instead.  Normally that's caught.

The ballot has marks at four races, but TEVS did not "find" the top-most mark, which, together with the other three marks, would have meant four votes in a race where the voter could vote for no more than three.  Instead of discarding all votes, TEVS therefore counted the Fox, Sommer, and Corley marks as valid votes.

The approach TEVS uses is to measure the average lightness or darkness of the vote target region and, where the darkness is borderline, check the region at the center of the target region.  If there's a dark pixel in the central zone (picture the central square of an overlaid tic-tac-toe board), TEVS calls it a vote.  Otherwise, it is flagged as suspicious, but not called a vote.  Because the voter's topmost "x" is a bit high, forming a circle in the vote target, TEVS did not read the vote.

I think I'd been thinking that by just checking the center, I wouldn't have to worry about cases where the printed dark rectangle marking the vote zone had bled into the vote region by a pixel or two. This is just a demonstration of how easily an attempt at optimizing for increased speed can let votes slip by.  

I believe the county's system will have rejected all four votes as an "overvote" condition, since there is no way to interpret one of the votes as less intentional than the others, and no indication of attempts by the voter to point out a correction.

Did Bergel Voters Vote For Arroyo? For Albin?

Because Humboldt County's elections office allows citizens to see results ballot by ballot, as all elections offices should, we in Humboldt can actually see how the group that voted for one choice voted for other choices.

I haven't had a chance yet to generate exact numbers, so I'm going to round all numbers here to the nearest ten.

Early Eureka Results

Since one race was close, I wanted to put the following up, although I will be working to further examine the results for the next few days. If anyone has questions, I can be contacted as mjtrac on gmail.  There will  probably be a few further adjustments, but they should only be by two or three votes.  Let me stress that this is basically confirmation of the elections office counts.  There will almost certainly be a few minor adjustments to these numbers.

Prior to checking for overvotes, TEVS counted the following in the close Eureka City Council race:

 count |  choice_text  
-------+----------------
  3364 | KIM_BERGEL     (adj. to 3361, see below) (county result 3359)
  3316 | MIKE_L_NEWMAN   (adj. to 3309, see below) (county result 3313)
    36 | Write-in   

The following images, on inspection, are overvotes that can be awarded to one or the other of the candidates, but were initially counted by TEVS (pre-overvote processing) as votes for both candidates:

66659, to Newman
17223, to Bergel
68415, to Bergel
23005, to Bergel
74843, to Bergel
16245, to Bergel
65199, to Bergel
65265, to Bergel
19767, to Newman
16717, to Newman

This reduces Bergel's count by 3, to 3361, and reduces Newman's count by 7, to 3309.  68415 is the most error-prone, where the mark is strong for one candidate, but the strong mark has a "NO" next to it and the weak mark has an arrow pointing to it. I considered it a vote for Bergel, but others may disagree, thinking that the NO and the arrow are meant to say NO to Bergel.

This is the sort of ballot that probably makes most elections officials hope the contests will never be decided by one vote.

Nov 2014 In Progress

Project volunteers finished scanning the last of Humboldt's 38K ballots on Friday, the day after the last set of ballots was provided to us by the county elections office.  Ballot images have been sorted into precincts for anyone who wants to count off the images, and I'm running software to generate an independent computer result, which ought to be available before Thanksgiving.

This election is notable for the narrow lead in one Eureka city council race, where the lead flipped between election night and the count of the last ballots.  This sort of close race highlights the importance of being certain that every vote was properly counted.  It also shines light on the oddness of election night "horse race" coverage, since, among other things, many ballots remain sealed inside envelopes for days after election night, awaiting the slow, ballot-by-ballot process of envelope signature validation.

It should also highlight the importance of allowing a citizens' group direct access to the cast ballots to allow them to create images, with non-elections equipment, digitally sign them, and confirm the elections office count.  Elections only work well when the loser is satisfied that the count was fair.

June 2014 results

The Transparency Project results for the June 2014 primary track the elections office results quite closely.

113,470 single side images were scanned from the cast ballots provided by the elections office.

Of these, all but 41 images were accepted by TEVS.  Without manually adding in these images, and without manually checking TEVS best guesses as to how to award overvotes, we have the following:

Measure N 951 yes to 797 no in the county's results; 951 to 798 in the TEVS results.

Measure M 244 yes to 127 no in both county and TEVS results.

District 4 Supervisor, county 2821 Bass to 2587 Kerrigan with 13 writeins; TEVS 2821 Bass to 2583 Kerrigan with 13 writeins.

District 5 Supervisor,county 3465 Sundberg to 2239 LaTour with 16 writeins; TEVS 3456 Sundberg to 2224 LaTour with 15 writeins.  (My suspicion is that county elections workers were able to identify and award various voter-corrected ballots that TEVS left as overvotes, simply seeing two dark marks but not understanding any cross-outs or pointing arrows.)

DA:  County count, then ETP
Dollison: 2154 vs. 2157
Klein 1396 vs. 1391
Firpo 7454 vs. 7420
Fleming 16487 vs. 16450
Write-in 51 vs. 53

As with the District 4 race, the discrepancies in the DA race are likely due to voter-corrected ballots where TEVS saw that two positions had been marked but where county workers saw corrections marked on the ballots and awarded votes.

If anyone were to want to invest further time, they could manually inspect each overvoted ballot and each vote TEVS rated as uncertain, and add in the votes on the 41 images not automatically processed, to get more exact figures from the data available to the transparency project.  However, these results track the county results closely enough in the close races that I see no need to go to that level of detail.

November 2013 results

The election of November 2013 had very low turnout, but a little bit of excitement was caused by a flip in the results of one close race between election night and the first post-election update.

We scanned the ballots and performed an independent count of the election using the most recent version of TEVS (ISO image and signature available via links at humetp.org) and the TEVS results were virtually identical to the county's second post-election update, varying by zero to three votes over each race, out of 14 thousand ballots cast.

Images are available from the Humboldt County Elections Office.

The complete image set, with a copy of the TEVS ISO, is also available on a 16 gig memory stick for $25.  Please email mjtrac at email provider gmail dot com if interested.