Take a county with 64,000 ballots and a race decided by 8 percent. To check the result by comparing computer records against paper, officials would need to pull about 80 ballots on average. To check it by polling a random sample, about 700. To check it batch by batch, about 13,000, spread over 26 batches. Those figures, for a 10 percent allowance of undetected wrong outcomes, show how a risk-limiting audit works: the size of the job depends on how close the race was.
A risk-limiting audit, or RLA, is a check run after an election to cap the chance that the reported winner is the wrong one. It has two parts. Voter-verified paper ballots are kept under lock until the check. Then staff examine a random sample by hand until the evidence meets a set risk limit. Colorado, for one, has used a 9 percent limit, which means the audit catches at least 91 percent of wrong winners.
The Brennan Center calls it the gold standard of post-election audits. It can only be done where every vote leaves a paper record the voter can check, and it is usually carried out in public.
How the sample grows
Wide margin, small sample. Close race, bigger one.
If the sample matches the official result, the outcome stands and the audit ends. If it turns up enough discrepancies to cast doubt, officials draw a larger sample and count again. That can continue until the sample confirms the original winner or a hand count of every ballot names a different one.
The numbers climb fast in tight contests. In a small city with 4,000 ballots, comparing ballots against computer records for a 2 percent margin takes 300 ballots. At a 0.1 percent margin it takes 3,000, or 50 to 100 staff hours. The other two methods would need a full hand count of all 4,000 at that margin, 70 to 130 staff hours. Margins under 0.1 percent turn up in one contest out of every 60 to 460.
Three ways to audit
In ballot comparison, election computers produce their reading of each ballot, and people check a random sample of those readings against the stored paper. An independent computer then retotals all the records. In ballot polling, people count a random sample and compare the percentages with the computer’s. In batch comparison, people hand-count every ballot in a random sample of batches, such as precincts, and confirm that the batch totals were added correctly.
Comparisons can run ballot by ballot or precinct by precinct, though the second costs more.
What it catches
An RLA looks for deliberate changes to ballots and tallies, and also for plain machine faults: software bugs, scanners with a blocked sensor, scanners that skip ballots. The sample is sized to have a good chance of catching a fleck of paper blocking one sensor on one scanner for a short time, or a bug or hack that flips votes in a single precinct, as long as enough ballots are affected to change the result.
It checks results, not procedures. And storage matters. If the sampled ballots show the boxes were not kept safely, the usual method cannot recover the right result, and researchers recommend a re-vote when enough ballots were held that way to change a winner. Seven Florida counties offer another route: they copy paper ballots the day after the vote on machines separate from the election machines.
Limits
RLAs have drawbacks. If any contest is close, the sample has to be a large share of all ballots. Computer totals are hard to check in public unless the records are released. No place audits every contest on the ballot, which leaves most local races unchecked even though millions of dollars in local spending and land-use decisions ride on them. Colorado picks contests with wider margins to avoid big samples. California’s rules for 2019 to 2021 required any RLA to cover all contests, and no election office chose to use one.
Maryland told its own planners that local boards could not budget for the work, because a very close margin could demand days of staff time.
