Start Here
Donate

Case 12 — Prince William County, 2020

12.1 — Introduction

On November 3, 2020, more than 220,000 residents of Prince William County, Virginia, cast ballots in the presidential election. Joe Biden won the county by more than 60,000 votes and carried Virginia by more than 450,000. No outcome was close.

But Prince William County reported the wrong numbers. Biden’s total was understated by 1,648 votes. Trump’s total was overstated by 2,327 votes. The net effect was a roughly 4,000-vote error in the margin of victory — in a county that is Virginia’s second-largest by population and one of the most diverse in the country.

The error went undetected for over a year. It was not caught by Virginia’s statewide risk-limiting audit. It was not caught by any routine post-election check. It was discovered by accident, by a new registrar responding to a public records request, who noticed that two precincts had reported identical vote totals — a statistical near-impossibility.

What followed was a criminal prosecution, a near-resignation, years of harassment directed at election staff, and ultimately the admission by the county that the errors were caused by technical incompatibility between the county’s vote reporting format and the state’s requirements. The criminal charges were dropped. The investigation found no fraud. And the county’s primary reform — publishing all ballot scanner tapes online — amounted to a partial implementation of the very concept that Actual Vote is built on.

12.2 — Background: The County and Its Reporting System

Prince William County is a wealthy suburb of Washington, D.C., centered on the city of Manassas. In 2020, it was Virginia’s second-largest county and among the most diverse in the country. It had been trending Democratic for years and was, by the time of the 2020 election, solidly blue at the presidential level.

Like most Virginia jurisdictions, Prince William County uses optical scan tabulators. Voters mark paper ballots, which are fed into scanners at the precinct. At the close of polls, each scanner prints a results tape — a receipt-like printout showing the vote totals for every contest on that machine. The vote data is also transmitted electronically to the county election office, which aggregates precinct-level results and reports them to the Virginia Department of Elections for certification.

Prince William County had an additional complication: split precincts. Several of the county’s polling places straddled the boundary between two congressional districts. In those precincts, the scanner counted all the votes together and printed a single results tape with aggregate totals. But the state’s reporting system required the presidential votes to be split by congressional district. The county’s voting system was not programmed to produce results in the format the state required.

This mismatch — between what the scanners produced and what the state expected — is the root cause of everything that followed.

12.3 — The Error

The Prince William County Office of Elections described the problem in a January 2024 statement: “The reporting errors were presumably a consequence of the results tapes not being programmed to a format that was compatible with state reporting requirements. Attempts to correct this issue appear to have created errors.”

In other words, the scanners counted correctly and printed correct tapes. But when the county tried to reformat the tape data to match the state’s reporting requirements — splitting aggregate precinct totals into separate congressional-district figures — the reformatting process introduced errors. Some votes were attributed to the wrong candidate. Some were double-counted. Some were lost.

Eric Olsen, who became director of elections in November 2021, later explained in a phone interview with the Associated Press that the majority of errors occurred in the split precincts. The county’s voting system did not break the presidential vote down by congressional district. The state system required it. The errors arose from the manual process of trying to conform the county data to the state requirements.

Separate from the split-precinct problem, there were also straightforward data entry mistakes. Olsen first discovered the reporting errors when he noticed that Precincts 607 and 608 displayed identical vote totals for both Biden and Trump. One matching number could be a coincidence. Two identical totals for two different candidates in adjacent precincts was, as Olsen put it, “an obvious typo” — someone had entered one precinct’s data into the other.

The aggregate effect across all races:

In the presidential race, Biden’s reported total of 142,863 should have been 144,511 — he was shorted 1,648 votes. Trump’s reported total of 81,222 should have been 78,895 — he received 2,327 more than he should have. The net swing in the margin of victory was approximately 3,975 votes.

The errors were not limited to the presidential race. In the U.S. Senate race, Democrat Mark Warner was shorted 1,589 votes and Republican Daniel Gade was shorted 107. In a U.S. House race, Republican Robert Wittman was shorted 293 votes. None of these discrepancies affected the outcome of any race — Biden won the county by more than 60,000 votes, Warner won statewide by more than 500,000, and Wittman won his district by more than 80,000.

As Olsen noted in his press release: “The reporting errors did not consistently favor one party or candidate but were likely due to a lack of proper planning, a difficult election environment, and human error.”

12.4 — Discovery

The errors were not caught by any official audit or review. They sat undetected in the official record for more than a year.

Eric Olsen replaced the previous registrar, Michele White, in November 2021. Shortly after taking the job, while responding to a public records request, he noticed the identical-precinct anomaly described above: Precincts 607 and 608 with matching vote totals for both candidates. He dug into the data and found the broader pattern of errors across the split precincts.

Olsen reported the discrepancies to the Virginia State Board of Elections.

It is worth pausing on the mechanism of discovery. The error was found because a new employee happened to look closely at the data in response to an outside request. Not because a routine check flagged it. Not because the statewide audit caught it. Not because anyone in the county noticed that precinct-level numbers did not add up. A public records request, a pair of fresh eyes, and an unlikely statistical coincidence were the only things standing between this error and permanent obscurity.

12.5 — The Statewide Audit That Missed It

Virginia conducted a risk-limiting audit of the 2020 election. The results, presented to the State Board of Elections in March 2021, showed greater than 99% statistical confidence in the reported outcome. The audit sampled 1,372 ballots across 122 of Virginia’s 133 localities.

The audit confirmed that the ballots were counted correctly. That is, the scanners read the ballots accurately and recorded the votes that voters intended to cast. This is what an RLA is designed to verify — counting-layer accuracy.

But Prince William County’s error was not a counting-layer problem. The scanners counted correctly. The tapes printed correctly. The error occurred afterward, when the correctly counted results were reformatted and entered into the state reporting system. This is the reporting layer — the chain of transmission between what the machines actually counted and what the public sees as the official result.

The Virginia RLA was not designed to catch reporting-layer errors, and it did not. This is not a criticism of the RLA — it did what it was designed to do. But it illustrates a gap in the post-election verification process. An audit that checks ballots against scanner totals will confirm that the scanners counted correctly. It will not tell you whether those scanner totals were faithfully transmitted to the official results. Those are two different questions, and they require two different kinds of verification.

12.6 — The Criminal Case

Olsen reported the discrepancies to the state board, which referred the matter to the Virginia Attorney General, Jason Miyares. Miyares’s office charged Olsen’s predecessor, Michele White, with felony corrupt conduct, making a false statement, and neglect of duty.

The case became the highest-profile prosecution undertaken by Miyares’s new Election Integrity Unit, which he announced in September 2022. Both Miyares and the Republican Party of Virginia pointed to Prince William County to make the case for why the unit was necessary.

But the case unraveled. In December 2023, the state dropped all felony charges, leaving a single misdemeanor for willful neglect of duty. In early January 2024, they dropped the misdemeanor as well. The stated reason: “inconsistencies in witness statements” — a government witness had changed his testimony during trial preparation.

White’s attorney, Zachary Stafford, said the charges had been disproven. He argued that the Electoral Board — not White — was the body that certified the incorrect results, and that White had been made a scapegoat: “The board certified incorrect results and they, and the attorney general’s office, attempted to assign blame to Ms. White for their mistakes.”

The investigation concluded with no finding of intentional misconduct or fraud. The root cause remained what the technical evidence suggested: tape format incompatibility, split-precinct complexity, and human error.

12.7 — The Aftermath

The consequences fell most heavily on the people who had nothing to do with the original mistake.

Olsen, who had discovered and reported the error, became a target. He was sued by local election-integrity activists seeking to decertify the 2022 election results (dismissed) and by the local Republican Party over the appointment of election officers (successful). He received harassing emails from as far away as Oregon. His health deteriorated — he announced during an October 2022 electoral board meeting that he would resign after the upcoming election because a serious heart condition was being aggravated by “debilitating stress.”

“If I’m dead next year, I won’t be a very good registrar anyway,” he told the board.

He ultimately reversed his decision and stayed on the job. His communications specialist, Thalia Simpson, summarized the situation plainly: “There was no fraud. It’s hard for people to realize we’re still humans.”

The error — a technical problem caused by formatting incompatibility — had been absorbed into the narrative of election denialism. Despite the investigation finding no fraud, despite the charges being dropped, and despite the error having benefited Trump rather than Biden, it continued to be cited as evidence that elections could not be trusted.

12.8 — The Reforms

The Prince William County Office of Elections implemented several changes to prevent similar errors:

The most significant reform is that the county now publishes all ballot scanner tapes on its website. As Olsen told The Guardian: “You can take that and compare it to the data on the state website, if you want, to make sure that we’re tabulating stuff correctly.”

Virginia’s most recent redistricting dramatically reduced the number of split precincts in the county, removing the structural problem that contributed most to the original errors. The county also implemented a more thorough canvass process, and the Virginia Department of Elections updated its reporting system with more rigorous standards for validating results.

12.9 — What Actual Vote Would Have Contributed

The Prince William County reporting error is a textbook case for Actual Vote. It illustrates the specific kind of failure that AV is designed to detect, and the specific gap in existing verification that allows such failures to persist undetected.

Immediate detection. AV volunteers photograph tabulator tapes at the close of polls. The tapes in Prince William County showed the correct vote totals — the scanners counted accurately. The error occurred only when those totals were reformatted for the state reporting system. An AV comparison of tape totals against the officially reported results would have flagged the approximately 3,975-vote presidential discrepancy as soon as the official results were released — within days of the election, not more than a year later.

The same comparison would have flagged the discrepancies in the Senate and House races. In every case, the tape totals (correct) would not have matched the reported totals (incorrect).

Filling the audit gap. Virginia’s risk-limiting audit checked counting-layer accuracy — whether the scanners read the ballots correctly — and confirmed that they did. But the RLA did not check the reporting layer — whether the scanner totals were faithfully transmitted to the official results. AV checks exactly this. The two methods are complementary: an RLA verifies that the machines counted right; AV verifies that the reported results match what the machines actually said.

In Prince William County, the counting was correct and the reporting was wrong. An RLA alone would never have caught the error. AV would have caught it immediately.

Independence. The county’s own reform — publishing scanner tapes online — validates the core AV insight: that making precinct-level results tapes publicly available is the right way to verify the reporting layer. But there is a meaningful difference between a county publishing its own tapes and independent volunteers photographing tapes at the precinct on election night. The county’s published tapes are government-controlled documents. AV photographs are independent records, captured before anyone has an opportunity to alter, correct, or reformat the data. In a case like this one — where the error was introduced precisely during the process of reformatting official data — that independence matters.

What AV Would Not Have Done

AV would not have determined whether the reporting error was intentional or accidental. It detects discrepancies between tape totals and official results; it does not explain why those discrepancies exist. The distinction between fraud and error requires investigation — but AV provides the evidence that triggers and informs that investigation.

AV would not have caught any errors in the counting layer — if the scanners themselves had miscounted before printing the tapes, AV would not have detected it. That is what the RLA is for.

The transparency benefit, independent of what AV would have found

The specific Prince William reporting error is one of countless localities in which AV evidence would have done useful work even if the reporting had been correct. Consider the 132 other Virginia localities in the same 2020 election whose official totals no one has ever checked against precinct-level tape evidence. In most of those localities, the official reporting was almost certainly correct. But “almost certainly” is not the same as “verifiably correct,” and the public has no independent basis for believing one over the other. An AV corpus from that election would have delivered transparency across the board: in the one locality where the reporting was wrong (Prince William), the independent evidence would have surfaced the error within days; in the 132 localities where the reporting was right, the independent evidence would have confirmed it was right, and would have stood as a publicly available rebuttal to any subsequent bad-faith accusation that Virginia’s 2020 totals could not be trusted. That second outcome — evidence that vindicates the official record against the possibility of future attack — is just as much a transparency benefit as the first.

The defensive frame: a counterfactual that did not happen

Prince William 2020 is also worth reading as a near-miss of the kind of bad-faith narrative that grew around its mirror case in Antrim. The error in Prince William ran in the opposite direction from Antrim’s: Trump was over-reported by 2,327 votes and Biden was shorted by 1,648, in a Republican-leaning Virginia county whose result the Biden campaign had no particular reason to scrutinize. The discrepancy sat in the certified record for more than a year before a new registrar discovered it by accident.

Imagine the counterfactual in which the discrepancy had become public during the unsettled days of November 2020 rather than in May 2022. A Republican-leaning county in a state Biden carried by ten points, reporting an unexpectedly large Trump margin in the presidential race, the U.S. Senate race, and a House contest — the same kind of multi-race anomaly that, in Antrim, fueled the “Dominion stole the election” narrative for years. The available story would have been mirror-symmetric: Republicans had inflated their numbers in a county where it would not change the statewide outcome but where the inflation would constitute evidence of Republican fraud at the precinct level. Some commentators on the left would have advanced rigorous versions of this argument; some would have advanced bad-faith versions of it. In neither case would the official rebuttal have been independent of the Republican-aligned local officials whose conduct was being challenged. The state RLA had already certified the election; the RLA had been the wrong instrument to detect this particular failure; the rebuttal would have rested on “trust the RLA” and “trust the county,” each of which would have been the structurally weak answer the moment the question became politicized.

What AV evidence would have contributed in that counterfactual is the same thing it would have contributed in Antrim, with the directional valence flipped. Citizen-held photographs of every Prince William precinct’s poll tape, captured at the close of polls on November 3, 2020, would have constituted a public record of what the scanners had read — independent of the registrar, independent of the state, independent of either party. Any subsequent narrative would have had to grapple with the photographs. The actual error would have been detectable from the photographs (the tapes would have shown numbers consistent with Biden carrying the county, not the inflated Trump numbers in the official report). The bad-faith narrative — Republicans stole Prince William — would have had to confront the same photographs and either verify or fall against them.

The two cases — Antrim and Prince William — together establish what no individual case can establish on its own: that reporting-layer errors run in whichever direction the misinterpretation happens to run, and that bad-faith narratives can grow around those errors from whichever political direction has rhetorical incentive to grow them. The defensive value of independent precinct evidence is bidirectional and structural. It does not depend on which side is doing the accusing or which side benefits from the underlying error. See Defending Against Bad-Faith Accusations for the broader frame.

12.10 — What We Know and What We Don’t

What we know:

Prince William County misreported its 2020 election results by approximately 3,975 votes in the presidential race, with smaller errors in the Senate and House races. The root cause was a formatting incompatibility between the county’s results tapes and the state’s reporting requirements, compounded by data entry errors. The errors were discovered more than a year after the election, by accident. A criminal prosecution of the former registrar ended with all charges dropped and no finding of fraud or intentional misconduct. The county has since published all scanner tapes online and implemented other reforms. Virginia’s statewide risk-limiting audit, which found greater than 99% confidence in the election outcome, was not designed to detect this type of error and did not catch it.

What we don’t know:

Whether other Virginia localities had similar formatting mismatches between their results tapes and the state reporting system that went undetected. The Virginia RLA confirmed counting accuracy statewide, but no systematic check of reporting-layer accuracy was performed across all 133 localities. The PWC error was found by accident. If comparable errors existed in smaller counties — where there might be no fresh-eyed registrar and no public records request to trigger a review — they would remain in the official record uncorrected.

12.11 — A Related Reporting-Layer Error: Henrico County, 2020

The same 2020 election produced a second Virginia reporting-layer failure that illustrates a different subtype of the same underlying vulnerability.

In Henrico County, a memory stick containing the results of 16,616 in-person absentee ballots was mislabeled as “provisional” rather than “absentee.” Because provisional ballots follow a separate processing workflow, the entire batch was excluded from election-night totals. The tabulators had counted correctly. The tapes had printed correctly. The votes simply did not reach the official count until the mislabeling was caught.

The mechanism is different from Prince William County’s. Prince William was an arithmetic aggregation failure: correct tape totals passed through a reformatting step that introduced new, incorrect numbers. Henrico was a categorical misrouting failure: correct tape totals attached to a container whose label sent the batch down the wrong pipeline. Different subtype, same root vulnerability — the absence of a precinct-level check that compares what the scanners produced against what the official results contain.

The distinction matters for the Actual Vote methodology. AV’s photograph of the poll tape is indifferent to which subtype of reporting-layer failure is at work. In Prince William, an AV comparison would have surfaced arithmetic drift between tape totals and state totals. In Henrico, it would have surfaced a missing batch — tape totals present on election night but absent from the official count. In both cases the discrepancy is visible at the earliest point where visibility is possible. In both cases the existing safeguards did not surface it on their own.

A note on sourcing: the description above relies on a secondary summary of the Henrico incident and has not been corroborated here against primary sources (Henrico County Office of Elections records, Virginia Department of Elections reports, or contemporaneous news coverage). The case is cited to establish the subtype, not to build an independent narrative. A standalone treatment — parallel to this one — would require that primary-source work before making specific claims about discovery, chronology, or remediation.

12.12 — Why It Matters

A roughly 4,000-vote reporting error persisted in the official record of a major Virginia county for more than a year. It was not caught by the county’s own certification process. It was not caught by the statewide risk-limiting audit. It was found only because a new registrar happened to look at the data while responding to a records request.

The error was ultimately benign in its effect on outcomes — no race was close enough for it to matter. But “the error didn’t change the winner” is not the same as “the error is acceptable.” A reporting system that silently misattributes thousands of votes and has no mechanism to catch the mistake is a system with a structural vulnerability, regardless of whether that vulnerability happens to affect any particular outcome.

Prince William County’s response — publishing all scanner tapes online — is an acknowledgment that the previous system had insufficient transparency at the precinct level. It is also, in essence, a government-run version of what Actual Vote does independently. The county recognized that the fix was to make the primary evidence — the tapes — available for anyone to check. AV provides the same evidence, captured independently, from a position outside the government’s chain of custody.

The Virginia RLA and AV are not competitors. They answer different questions. The RLA asks: did the scanners count the ballots correctly? AV asks: do the official results match what the scanners actually reported? In Prince William County, the answer to the first question was yes, and the answer to the second was no. A system that only asks the first question will miss every error of the second kind.

12.13 — Further Reading