If we talk about this any more, let's take it off-list.
On Wed, Apr 19, 2006 at 10:24:11AM -0700, James Richard Tyrer wrote:
> Dieter wrote:
> >In message <[EMAIL PROTECTED]>, Jack Carroll writes:
> >
> >> The one real weakness with the present system, optical scanning of
> >>hand-marked ballots, is that nobody orders a manual recount unless they
> >>suspect an error, which they usually don't unless the election is close.
> >>An
> >>internal fault could produce a large error as easily as a small one.
>
> The machines are supposed to check themselves. A machine that reads the
> ballot twice would be much more accurate even if it had minor errors.
Supposedly, but we don't really know. They're designed and marketed
as off-the-shelf commercial products, and the public and the state officials
don't get to go through the designs with a fine-tooth comb to prove 100%
fault coverage. Trade secrets is the justification.
> >Count all votes twice, once with Brand-A machine, once with Brand-B
> >machine, which is required to be significantly different from Brand-A.
> >If the counts differ, run them through again. Still different?
> >Hand recount.
Right now, there's only one brand that can be programmed to read the
new NH columnar ballot, and it's already been shown to be hackable. Not to
mention, built by a company involved in multiple scandals. But this is
along the same lines as one of my suggestions, to allow anyone who has the
right to station poll inspectors and request a recount to make an unofficial
machine count with their own unapproved equipment. That wouldn't be
certifible as an official result, but it would provide real data for a
rational decision on when to recount.
There's another problem, though. Voters can and sometimes do mark
ballots in a way that a machine can't understand, but a human can. Circling
the name instead of filling in the oval. Writing across the top "Straight
Republican Ticket". Best answer I know of is to have the machine set any
ballot aside that has marks outside the ovals, for a human to count.
>
> I wonder if they test the machines with a validation stack of ballots.
Yes, but that only picks up a limited number of solid faults. They
do it in NH to check the election configuration data, not the firmware and
hardware.
About 5 years ago I was assigned to design an FPGA emulation of an
electric meter chip that was under development. The idea was to verify all
the internal logic before laying it out and fabbing prototypes. Once I got
the emulation board working, the project leader had me write a debugger on a
PC that could send commands to it, and exercise it. Next, he asked me to
write a little Basic program to run on a PC, and generate files of debugger
commands with pseudo-random number data, calculate the results the emulator
should get, and compare the results. He figured a couple thousand random
test cases would be enough. OK, I did that. It looked OK, but it only took
a couple of seconds to run. So I figured, OK, I just spent 3 months
designing this thing and getting it built, why not run it a little more? So
I upped it to 10,000 test cases. Still looked good, but that took under a
minute. Well, how about 100,000 cases? WHOOPS! 3 arithmetic errors! Ran
it again with the same pseudo-random numbers. Same result, on the same 3
cases. Reported that, went on to run 2,000,000 cases. Took an hour. It
averaged one error for every 30,000 cases. The project leader examined the
particular cases, and tracked it down to a carry error in one of the lower
bits of a 48-bit multiply.
That's why I don't have a lot of confidence in black-box testing
with a limited number of cases. For safety-critical avionics controls,
RTCA/DO-178 requires analyzing every path through the logic, and proving
fault coverage of every path and every decision alternative. That isn't
even possible with an off-the-shelf microprocessor, because only the
manufacturer has accesss to its exact logic design.
> >Hand recounts of a few randomly selected precincts.
Won't pick up low-probabability logic errors or random malfunctions,
not to mention the dreaded intermittents.
> Problem there is that the humans probably can't do any better than the
> machine.
>
> In theory, the humans can never get it exactly correct, but the machine
> will either have at least 3 '9's (99.9% correct) or it will make large
> errors due to a malfunction. Problem is that there is the possibility
> of an intermittent malfunction of the machine. Intermittent
> malfunctions drive repairmen nuts.
I attended a meeting the NH Secretary of State hosted last night,
with some local election officials and other interested people. I learned a
lot.
NH has had 300 years to get hand counting right; about half the
state has never used machines. Exact procedures differ from one polling
place to another, but there's quite a lot of cross checking. A typical team
has 3 people on it, with one reading, one tallying, and one watching both.
After each batch of 50, the votes, non-votes, and defective ballots for each
choice get added up, and if they don't total 50, the batch is counted again,
with the team members switching seats. There's an election officials'
procedures handbook, which explains various methods that have stood the test
of time. In some towns, each batch is counted by two teams, and the tally
sheets are compared before adding up the results. There are always two
clerks adding up the tally sheets independently, and if they use
calculators, they don't both use the same one.
The one real problem that comes up is the ballot that makes the
counters go "What the #$^@&% did THIS voter mean???!" The Secretary told the
story of one ballot on a recount that looked like it was full of overvotes.
But, on a one-seat four-candidate race three boxes were filled in. On a
two-seat six-candidate race, four were filled in. On some races, all the
boxes were filled in. Same pattern throughout the ballot. What the
officials finally decided was that the voter had blackened all the boxes for
candidates and measures he _didn't_ intend to vote for. Not a certainty,
but sometimes certainty isn't possible. A machine would have accepted the
few races where only one box was filled in (probably the opposite of the
voter's intent), and rejected the rest as overvotes.
In fact, recounts after machine votes generally add some votes to
each side of a choice, which the machine couldn't interpret or even see. It
rarely changes an election outcome, but it does happen.
> The scanners they use in Arizona are at the polls and they are very
> picky about accepting ballots -- they kick them out if they can't read
> them correctly. You insert the ballot and wait a few seconds while it
> decides to accept it or beep and it comes back out just like a bad
> dollar bill in a vending machine.
What then? Is the ballot set aside for hand counting, or does the
voter get a fresh ballot and instructions on how to mark it correctly?
>
> --
> JRT
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)
