Scanning ballots in Boulder County

[Paul Walmsley <paul@xxxxxxxxxxx>]

Hello everyone,

As some of you know, I decided to sign up as a election judge for this
November's election, helping to operate some of the County's vote-counting
gear.  Boulder County is using an optical scan paper ballot system made by
Hart InterCivic.  My first day scanning ballots was this past Friday, and
I figured that perhaps some of the people on this list might be interested
in hearing a description of how the process works.  It's fairly technical
and will probably bore most people to tears: you've been warned!  I've
also copyrighted this narrative: please see the end of the article for
license details.

Note that below I use the term "ballot image."  Recently, this term has
been corrupted.  The common-sense interpretation of this term is "a
scanned facsimile of a physical ballot marked by a voter," which is how I
use it below.


The Hardware
------------

I believe Boulder County has eight ballot scanning stations that can be
used for scanning ballots this year.  (The County has at least one other
scanner used for scanning absentee ballot envelopes - but I don't have any
experience with that system, so I can't write about it.)  These are 
all located at the Clerk and Recorder's office on 33rd St.  

Each of the scanning stations is marked with a letter, starting with "A",
which is then labeled on all of the computer hardware at that station.  
Each station also has its own particular color of printer paper that it
prints its reports on.

A ballot scanning station consists of a Kodak i830 scanner, three Dell
computers running Windows 2000, an PCMCIA flash card reader, a HP Laserjet
1012 printer, a video projector, and a KVM switch to switch the single LCD
monitor, keyboard, and mouse between the three computers.  All of this
gear (with the exception of the scanner, which is quite large) is set up
on a 7 or 8 foot high computer desk arrangement.  There's also a
collection of UPSes behind the station - there were three of them on the
scanning station I was at.  The three computers are networked to each
other.

One of the three computers is labeled "BN" - for "Ballot Now" - and the
other two are marked "BNIP" - for "Ballot Now Image Processor," I think.
Ballot Now is the name of the Hart InterCivic paper ballot processing
software.  It handles the ballot scanning process, and is able to generate
certain kinds of reports on the scanned ballots.  I would hazard an
educated guess that the BN machine (as opposed to the BNIP machines) is
the machine which reads the barcodes on the scanned pages to determine
whether to accept or reject a particular ballot.  More on this process 
below.

The two BNIP machines are slave machines that do the ballot interpretation
work.  When the scanner operator saves a batch of ballots after the
scanning process, the two BNIP machines start converting the scanned BMP
image files to Cast Vote Records (CVRs).  As best I can tell, CVRs are
data structures containing the machine's interpretation of each race on
each ballot, i.e., race 1, Badnarik; race 2, Salazar; etc.  At some point,
these Cast Vote Records are stored on a PCMCIA flash card connected to the
BN machine, but I don't know if this happens before or after the 
resolution process.


Resolution and Tallying
-----------------------

At some point after the ballots are scanned, the ballots must be resolved, 
and later, tallied.  

I haven't seen a ballot resolution process on live ballots yet.  Ballot
resolution is the process of reviewing and manually interpreting votes 
that the software was unable to classify automatically, for whatever 
reason.  So, for example, say that Ballot Now is set to require manual 
resolution of overvotes -- e.g., a vote for two candidates in a 
vote-for-one race.  As I understand the process, the County will project 
an image of the ballot up on the wall, and a team of election judges 
will doublecheck to ensure that it really was an overvote rather than a 
machine misinterpretation.  

(My guess, by the way, is that the ballots scanned on a certain station
must be resolved on that station.  The ballot images would be too large to
fit on the flash card, and to the best of my understanding, the scanning
stations are not networked together.  So in other words you couldn't 
resolve ballots from station "A" on station "B".)

The resolution changes are written to the PCMCIA flash card at some point.  
Then on election night, elections officials will take the flash cards over
to the "Tally" system.  "Tally" is actually the name of the Hart software.  
Officials will insert one flash card at a time into this machine, which is
a Dell computer with an outboard flash card reader and two printers
attached: one laser printer for report data, and one dot-matrix printer
for audit log data.  This machine will load all of the CVR data from the
flash cards, and then add up all the votes and generate a vote tabulation 
report.
 

The Ballot Scanning Process
---------------------------

The process that I actually was involved with on Friday was the ballot 
scanning process - the first step in the chain.  Here's how the process 
worked.

The cast ballots are stored in red transfer cases about the size of a
ballot and about a foot deep.  A group of election workers which I will
call the "ballot sorting team" would unseal these cases and page through
each stack to ensure that the ballot pages were in order and that the
ballots were all facing the same direction.  This is because the Hart
scanning software would reject any multipage ballot where the pages came
in out of order - e.g., page two before page one - which would mean 
those ballots would have to be rescanned later.  More on this later.

Once the ballot sorters were done with a red case, it would be transferred
to the scanning station responsible for the particular early voting
location.  The scanning station that I was at -- station G -- was assigned
to run early voting ballots from the Clerk & Recorder's Office ("33rd
St.").

We'd take a portion of the ballot stack out of the case - a few hundred
sheets - and load these into the scanner as a "batch."  (These batches are
automatically numbered by the Ballot Now software in ascending order,
starting with "1".  The batch numbers are only unique to that scanning
station - so, for example, station A would also start numbering at "1".)  
We'd type in a short descriptive note with the batch, listing the location
that the ballots were cast, the date they were cast, and the scanning
station letter.  We'd then tell the scanner to start scanning.

Once the scanner itself finished, we'd need to wait for a few seconds
while the computer caught up to it.  The computer would then display three 
numbers on the screen: the number of images processed, the number of 
ballots accepted, and the number of images rejected.  (An "image" is one 
side of one page of a ballot.  So a four-sided ballot would consist of 
four images - both sides of two pages.)

We would then click on a button marked "Review."  This would generate a
report which listed every ballot image processed, its sequence number and
image number, and whether the image was "Accepted" or not.  If not, the
report would list a short explanation why -- for example, it might read
"Page out of order," or "Duplicate ballot," or something along the lines
of "Unknown form."  We would "print" these reports as PDFs via the Adobe
Acrobat Distiller software, which would then open the PDF.  We then
printed the report onto the attached laser printer.

Once we had the report, we'd determine whether any images were rejected.  
If so, we had to go through the stack of scanned ballots and manually
search for the sequence numbers in question and pull those ballots.  The 
ballots would then go into either a "rescan" bin to be processed later, or 
into a "duplication board" bin destined for the dup board.

We would then fill out a short manual report on the batch totals,
initial it, and load the scanned ballots, the computer report, and the
batch report into a blue transfer case.  The blue transfer cases were
probably a little over twice as high as a single red transfer case.  When
the blue transfer case filled, it would be sealed with a wire and plastic
seal with a serial number stamped on it, and the blue case would be placed
in a storage room.  We'd also write the seal number of the blue transfer
case onto a transfer log page at the scanning station.

Finally we'd click "Save."  This would save the batch, and start the BNIP
machines processing the ballot images.  The machine's batch count would
increment, and we could then go on to scan the next batch of ballots.

....

More to come later.


- Paul

Copyright (C) 2004 Paul Walmsley.  Distribution of this message for
non-commercial purposes is permitted as long as the message is distributed
in full without modification.  Distribution for any commercial purpose
requires the author's express written permission.