Risk-Limiting Audits
A risk-limiting audit (RLA) examines a random sample of paper ballots, comparing them to the machine count to ensure that the winner actually won. Conducting them is as simple as pulling and reviewing a random assortment of paper ballots compared to the computer tabulation. The closer the results of the race, the more batches and ultimately, votes, are used to perform the RLA.
Arizona
Methodology
In Arizona, each county must conduct a hand count of 1% of early ballots and 2% of Election Day ballots.
Maricopa County's RLA consisted of 26 randomized batches of nearly 200 votes each, totalling 5,130 votes from an Early Voting population of 1,805,077.
This was followed by five more batches taken on Election Day from a population of 249,838.
Discrepancies in audit reports
In Arizona's largest counties (Maricopa and Pima), there were either misreported numbers or process deviations in audit reports for Prop 138 and 139, where the number of reviewed ballots exceeded the maximum possible batch sizes (287/200 max, 778/300 max).
Mismatch in extrapolated population
Early voting. In this random sample of Maricopa's Early Voting ballots, Harris has 53.1% while Trump has 46.3%. Extrapolated to the Early Voting population, Harris gets 959k votes, Trump 836k.
Actual reported percentages are Harris with 48.9% and Trump with 49.7%. In the random sample, Harris outperforms her reported number by 4.2% and Trump underperforms by 3.4%.
Election Day. In this random sample of Election Day ballots, Harris has 39.4% while Trump has 59.7%. Extrapolated to the Election Day population, Harris gets 98k votes, Trump 149k.
Senate totals from Election Day are also all multiples of 5, which again, while possible, improbable (source)
Extrapolated totals
Harris: 959k + 98k = 1,057k
Trump: 836k + 149k = 985k
Reported results
Harris: 980k
Trump: 1,052k
What's weird
Maricopa's RLA points us to the nearly the exact values of reported results but in reverse. The same pattern shows up in Pima County.
Source | Maricopa RLA | Pima RLA
Party Representative Mismatch
For early voting, the RLA requires at least 25 ballot batches to be reviewed, which are randomly chosen from a pile of at least fifty batches. Each batch is then selected to be hand-counted by a representative of either party.
The chart below shows the ratio of batch results (dark) and the ratio of a representative choosing a batch with their party majority (light). Take a look at how tall the dark and light bars are to each other.
In 2012, from 30 batches, the spread between batch result and same party representative pick was 11% for both parties.
In 2016, from 25 batches, the spread was 3%
In 2020, from 26 batches, the spread was 2%
In 2024, from 26 batches, the spread was 26%
There were 7.5 Trump Majority (29%) and 18.5 Harris Majority (71%) batches
Of eleven times when a representative picked a batch with their party majority, the Republican Rep picked 6 (55%) while the Democrat Rep picked 5 (45%)
What's weird
If the batches selected are truly at random, then the final ratio of the representative selecting a batch with their party majority should be reflective of this statistic. Instead, contrary to the RLA's actual results, we saw Republican representatives were much more likely to choose a batch where Trump won, and Democrat representatives less likely to choose a batch that Harris won, as well as a strong skew towards Harris overall.
Source | Maricopa RLA
Georgia
Methodology
Georgia's RLA consists of a manual inspection of random samples of paper ballots.
For 2024, county officials audited 442 batches of ballots, or about 14% of all ballots. Of the 442 batches, 381 (86.1%) had no deviation from the original candidate vote totals. The other 61 fell within an expected margin of error for a hand count.
Discrepancy Analysis
This analysis looked at the differences between reported and audited counts per batch and whether each discrepancy favored the winner (Trump), then determined how likely those discrepancies were to occur by chance alone.
It reveals that advanced in-person votes and votes scanned in local precincts show statistically significant discrepancies, warranting further investigation to determine their cause and potential impact.
Advanced In-Person Votes: Discrepancies affecting the margin for advanced in-person voters show a p-value of 3.85%. This indicates a meaningful deviation from what would be expected under random chance.
Imagecast Precinct Votes: Discrepancies affecting the margin for votes scanned in local precincts (Imagecast Precinct) show a p-value of 0.77%. This is even more concerning given the decentralized nature of local precinct operations and their potential vulnerabilities.
Source | GA RLA
Pennsylvania
Methodology
Pennsylvania's RLA consists of both statewide and county-level RLAs.
For the state, officials roll a die to choose randomized batches from random counties. In total, 55 batches of ballots in 32 counties were randomly chosen to be audited.
For counties, officials review a random sample of at least 2% of all ballots cast, or 2,000 ballots, whichever is fewer.
Non-representative samples
Philadelphia County represents roughly 10% of all presidential and statewide votes in Pennsylvania, but only accounted for 0.5% of the RLA sample. All the information represented by Philadelphia County was based on only 183 ballots – the third smallest sampling of all 55 batches.
Contrary to state guidelines, less than half of all counties (32/67) were represented in the sample
State guidelines also note that top-of-ballot eligible contests should always be a target contest of the RLA, but the 2024 RLA only looked at the State Treasurer race
Source | PA RLA