Case 13 — Antrim County, 2020

13.1 — Introduction

At 4:09 a.m. on November 4, 2020, Antrim County, Michigan, posted its unofficial results for the presidential election. The screenshot that appeared on social media within minutes showed Joe Biden beating Donald Trump in Antrim by more than 3,000 votes — 7,769 to 4,509 — in a rural, heavily Republican county of roughly 22,000 registered voters. Antrim had voted for Trump by nearly thirty points in 2016. The numbers were impossible.

The numbers were also wrong. The paper ballots themselves had been correctly marked and correctly scanned. The scanner memory cards correctly recorded what voters had done. The problem was that the county’s Election Management System, when it read those memory cards to aggregate the results, interpreted them using an updated ballot definition that the scanners themselves had not been updated to use. The result was a cascade of misattributed votes: Trump’s votes read as Biden’s, Jorgensen’s votes read as Trump’s, and similar shifts throughout every down-ballot race on every alphabetically-later ballot design in the county.

Within forty-eight hours, Antrim posted corrected results showing Trump winning the county by nearly 4,000 votes — roughly the margin any observer of Antrim politics would have expected. Six weeks later, a publicly-streamed hand count ordered by the state confirmed the corrected numbers to within twelve votes. Seven months later, a GOP-led Michigan Senate Oversight Committee called the conspiracy theories surrounding the county “a complete waste of time to consider further”. Seventeen months later, a unanimous Michigan Court of Appeals panel — including two appointees of Republican Governor Rick Snyder — dismissed the lawsuit built around the anomaly. In 2022, a rigorous forensic analysis of the incident won best-paper recognition at the USENIX Security Symposium.

None of that stopped the original 4 a.m. screenshot from becoming, by some measures, the single most consequential piece of election misinformation in modern American history. It was cited in dozens of post-election lawsuits, featured in materials circulated before January 6, 2021, and eventually contributed to a $787.5 million defamation settlement between Fox News and Dominion Voting Systems in 2023.

This is a case study about the gap between two clocks. One clock measures how long it takes for a reporting-layer error to become visible as a wrong public number. In Antrim, that clock ran for roughly forty-eight hours. The other clock measures how long it takes for the technical explanation of the wrong number to be publicly established, believed, and absorbed into the conversation. That clock is still running. The purpose of Actual Vote is to collapse the first clock to near-zero — to make the tape-vs.-report comparison available in real time, independently, so that the wrong number cannot establish itself as an uncontested fact while the forensic explanation is still being written.

Antrim 2020 is also the mirror-image companion to the Prince William County, Virginia, case (see case study 03). Both are textbook reporting-layer failures in the Actual Vote sense. In Prince William, the initial error favored Trump and the correction favored Biden. In Antrim, the initial error favored Biden and the correction favored Trump. The two cases illustrate a point that matters more than any single incident: reporting-layer errors are technical artifacts. They run in whichever direction the misinterpretation happens to run. Independent verification is not a partisan tool. It is a public good.

13.2 — Background: The County and Its Election System

Antrim County sits in the northwest of Michigan’s Lower Peninsula, between Traverse City and the tip of the mitten. It is rural, forested, and sparsely populated. Its eighteen townships together encompass 22,082 registered voters. It has voted Republican in every presidential election since 1972.

Antrim uses the Dominion Democracy Suite voting system, version 5.5. The architecture has three relevant components for this case:

The ImageCast Precinct (ICP) optical scanners at each polling place read hand-marked paper ballots, record the votes digitally on a memory card, and print a paper poll tape at close of polls summarizing the tabulated results for every contest.
The Election Management System (EMS) is a central server at the county election office that reads the memory cards from the precincts and aggregates the precinct-level totals into county-wide reports.
The election definition is a data file that tells both the scanners and the EMS how to interpret ballots — which candidate corresponds to which position on which ballot design. It is produced using Dominion’s Election Event Designer (EED).

The critical technical detail — the one that made Antrim’s 2020 incident possible — is that Dominion’s election definitions assign each candidate a sequential numeric identifier called a ChoiceManifestation ID. These IDs are assigned in alphabetical order across every contest on every ballot style in the county, in one continuous sequence. If a new write-in blank is added to one contest, every ID after that point shifts by one. A ballot marked under one version of the definition cannot be correctly interpreted by an EMS running a different version — the IDs no longer line up.

For most elections, this is invisible. The election definition is finalized once, distributed to both the scanners and the EMS, and never changes. Antrim’s 2020 election was not most elections.

13.3 — The Error

In mid-October 2020, Antrim County identified three ballot design errors that required correction in three specific townships:

Central Lake Precinct 1V. The school board contest needed to be changed from “Ellsworth Schools” (vote-for-two) to “Central Lake Schools” (vote-for-three). This involved adding a write-in blank.
Mancelona Precinct 1V. The Village Trustee contest needed candidate Eugene K. Kerr added, and the rule changed from vote-for-three to vote-for-two.
Warner Precinct 1BF. A new contest — the Boyne Falls Public Schools Sinking Fund Millage Proposal — needed to be added to the ballot.

These changes were regenerated in the Election Event Designer and loaded into the EMS. But the county’s election staff distributed the updated memory cards to only one of the three affected townships: Mancelona. Central Lake’s scanner and Warner’s scanner entered election day running the initial election definition, while the EMS ran the revised one.

The consequence was dictated by the arithmetic of the ChoiceManifestation IDs. Adding the write-in blank in Central Lake’s school board contest shifted every subsequently-numbered candidate ID by one. Alphabetically, Central Lake comes early in the township list. The shift therefore cascaded through every subsequent ballot design in the county. On election night, sixteen of eighteen scanners read memory cards under the initial definition’s numbering scheme; the EMS then loaded those cards and interpreted the IDs under the revised numbering scheme. Every ID was off by one.

The result, as documented by J. Alex Halderman’s forensic analysis, was systematic misattribution of votes in the presidential contest (and in every down-ballot contest):

Votes cast for Trump (initial ID 970) were counted as votes for Biden (revised ID 970).
Votes cast for Libertarian Party candidate Jo Jorgensen (initial ID 975) were counted as votes for Trump.
The same shift affected the straight-party ticket option, where selecting “Republican” was counted as selecting “Democratic.”
Most votes originally cast for Biden were counted as no selection at all.

A parallel problem affected three additional memory cards on election night. Warner’s card failed to load into the EMS because the revised definition included a new contest (the Boyne Falls Sinking Fund Millage) that did not exist on the card. Two other cards — Elk Rapids 1 and Milton 1 — were mistakenly “re-zeroed” by poll workers after the poll tapes were printed, discarding the digital results and forcing manual entry of the tape totals. County staff entered the three sets of manual results and posted the unofficial county-wide total at 4:09 a.m. on November 4.

That report — the one that produced the viral screenshot — showed Biden leading Trump 7,769 to 4,509 in Antrim County. The correct totals, which would emerge over the following days, were Trump 9,748 to Biden 5,960. The initial report was wrong by a margin of over 7,000 votes in the presidential race, and contained similarly impossible numbers in contests from U.S. Senate to State Representative.

A second layer of error then compounded the first. The county’s corrections on November 5 were still wrong: three precincts (Custer, Echo, Elk Rapids AV) had both the manually entered results and the incorrect memory-card results summed together, because the operator neglected to mark the memory-card results as “rejected” after adding the manual entries. The county issued amended corrections on November 6, and the County Board of Canvassers certified the results later that day.

Even the certified results were not yet clean. Michigan law requires county canvassers to compare certified results to the individual scanner poll tapes. Antrim’s canvassers did not correctly complete this comparison. Over the following two weeks, staff identified data entry errors affecting approximately 2.6% of votes county-wide — contests omitted, candidate totals transposed, typographical errors in specific candidates’ totals — spread across Kearney, Chestonia, Custer, Milton, Star, and Elk Rapids precincts. The county restated its certified results on November 16 (first amended) and again on November 21 (second amended), which remain the final results.

The scanner poll tapes themselves — the receipts printed at each precinct at the close of polls — correctly reflected what the scanners had read. The errors lived in the translation from tape to report.

13.4 — The Numbers

The county published five separate sets of results in a three-week window. They tell the story by themselves.

Report	Date	Biden	Trump	Biden margin
Initial unofficial	Nov. 4, 4:09 a.m.	7,769	4,509	+3,260
Partial corrected	Nov. 5	2,839	9,661	−6,822
Amended unofficial	Nov. 6	5,959	9,748	−3,789
Certified, 1st amended	Nov. 16	5,960	9,748	−3,788
Certified, 2nd amended (final)	Nov. 21	5,960	9,748	−3,788
State hand count	Dec. 17, 2020	5,959	9,759	−3,800

The gap between the initial unofficial report and the hand count is more than 7,000 votes in a county where roughly 16,000 people voted. The gap between the final certified result and the publicly-conducted hand count is twelve votes — a 0.1% difference attributable entirely to routine marginal-mark interpretation, the same kind of small discrepancy that appears between any machine count and any hand count.

The machine count was essentially accurate at the level of what voters had actually marked. The initial report was essentially a work of fiction produced by an EMS reading memory cards it could not correctly interpret. And the correction from fiction to fact took the public record seventeen days.

13.5 — The Narrative Gets Ahead of the Facts

By the time Antrim posted its first corrected results on November 5, the 4:09 a.m. screenshot from the previous morning had already traveled. It showed up in tweets, in pro-Trump forums, in the early drafts of what would become dozens of post-election lawsuits. The narrative it fed was simple and sticky: a Dominion voting system, in a Republican county, had somehow produced a Biden victory, and only after “they” realized people were watching did the numbers mysteriously “correct” themselves.

On December 6, 2020, 13th Circuit Judge Kevin Elsenheimer — himself a former Republican state legislator — granted a motion by attorney Matthew DePerno and Antrim resident William Bailey to conduct an outside inspection of Antrim’s election equipment. DePerno brought in Allied Security Operations Group (ASOG), led by Russell Ramsland, Jr. ASOG imaged the EMS and scanner memory cards.

On December 13, 2020, ASOG released the “Antrim Michigan Forensics Report,” authored by Ramsland. Its central conclusion was direct: “The Dominion Voting System is intentionally and purposefully designed with inherent errors to create systemic fraud and influence election results.” The report claimed a 68.05% “error rate” in Central Lake, a “staggering number” of adjudicated votes, evidence of unauthorized vote reversal, and parallels to voting machine fraud in Venezuela.

The ASOG report became the factual core of the “stolen election” theory. It was cited in the Sidney Powell “Kraken” filings, in the Texas v. Pennsylvania Supreme Court filing, in the Trump campaign’s Michigan election-contest filings, and in materials used to brief lawmakers in the weeks before January 6. DePerno raised approximately $400,000 for an “Election Fraud Defense Fund” centered on Antrim.

Meanwhile, on November 7, the Michigan Department of State had already published a fact-check titled “Isolated User Error in Antrim County Does Not Affect Election Results, Has No Impact on Other Counties or States”. On December 17, the State Bureau of Elections conducted a full hand count of the Antrim presidential contest. The count was open to the public and streamed online. The result: Trump 9,759, Biden 5,959. Net change from the certified results: twelve votes.

None of this kept pace with the screenshot.

13.6 — The Facts Catch Up (Slowly)

Over the following two years, every institutional mechanism that existed to review the Antrim claims reached the same conclusion. The conclusion was always the same. The time it took was always long.

May 18, 2021. Judge Elsenheimer dismisses Bailey v. Antrim County, citing the completed hand count and statewide audits. He notes that the statute Bailey invokes does not permit private citizens to conduct their own audits under terms of their choosing.
June 23, 2021. The Michigan Senate Oversight Committee — Republican-controlled, chaired by Sen. Ed McBroom (R-Vulcan) — releases its eight-month report. On Antrim specifically: “All compelling theories that sprang forth from the rumors surrounding Antrim County are diminished so significantly as for it to be a complete waste of time to consider them further.” The committee recommends that Michigan Attorney General Dana Nessel investigate “those who have been utilizing misleading and false information about Antrim County to raise money or publicity for their own ends.”
March 26, 2021. J. Alex Halderman, a professor of computer science at the University of Michigan and co-chair of Michigan’s Election Security Advisory Commission, files a forensic report under penalty of perjury as part of the Bailey litigation. Fifty-plus pages. Every anomaly in the Antrim election accounted for, with references to primary documents. Every substantive claim in the ASOG report rebutted point by point.
April 21, 2022. The Michigan Court of Appeals unanimously (3-0) affirms the dismissal of Bailey v. Antrim County. Two of the three judges are appointees of Republican Governor Rick Snyder. The court finds that “plaintiff merely raised a series of questions about the election without making any specific factual allegations as required.”
2022. Halderman’s technical paper on the Antrim incident is presented at the USENIX Security Symposium, one of the premier venues for academic computer security research. It receives best-paper recognition.
December 9, 2022. The Michigan Supreme Court declines further review, ending the litigation.
2023. Halderman’s state-court forensic report is publicly released, after years under seal.

Every one of these institutions agreed on the same answer. The initial error was human. The corrected results were accurate. The ASOG theory was technically wrong. The machine count had been independently confirmed by a hand count of the actual paper ballots. The litigation had no factual foundation.

The screenshot, by that point, had already traveled around the world several times.

13.7 — What the Forensic Record Actually Shows

Halderman’s forensic report is worth summarizing briefly, because it is the primary record of what happened and it speaks directly to claims that remain in circulation. What follows is a distillation of Section 5 of the state-court report. Readers who want the full rebuttal should consult the report directly.

The claim that ballots were “adjudicated” at scale. Dominion’s adjudication feature — which allows an election worker to review a scanned ballot image and manually correct the scanner’s interpretation — was not enabled in Antrim County. The adjudication software component was not installed. The scanners were not configured to store ballot images (a necessary precondition). The EMS contains no adjudication log entries for any election in any period. Ramsland’s claim of a “staggering number” of adjudicated votes refers to an event that did not occur.

The claim of a 68.05% error rate in Central Lake. This number was calculated from scanner log lines, not from ballots or votes. Scanning a single ballot produces anywhere from two to dozens of log lines, including routine warnings about paper misfeeds, pre-election test entries, and diagnostic messages. The fraction of log lines that are “errors” bears no relationship to the fraction of ballots or votes that were incorrectly processed. The 68.05% figure is mathematically meaningless as an error rate.

The claim that “ballot reversed” entries in the log indicate ballot rejection. The log message “Ballot has been reversed” refers to the paper-feeding mechanism physically returning a misfed ballot to the voter — analogous to a vending machine returning a dollar bill that was inserted incorrectly. The ballot is then typically fed in again successfully a few seconds later, as the logs show. It has nothing to do with rejection or adjudication.

The claim that the Antrim EMS was connected to the internet. The EMS system event log shows no evidence of any network connectivity. Antrim did not purchase Dominion’s wireless transmission functionality and did not use it. Memory cards were physically transported between precincts and the county office.

The claim that “software updates” were applied between scans. The ballot design revisions were not software updates. They were election definition data files, regenerated in the Election Event Designer. The scanner software and firmware were unchanged. Ramsland conflates a data-file change with a software change.

Halderman also confirms that the ASOG report identified some real vulnerabilities in Antrim’s election infrastructure: missing Windows security updates on the EMS, weak authentication and access-control mechanisms, and an unencrypted EMS hard disk. These are genuine problems that should be addressed. None of them were exploited in 2020. The anomalies that occurred in Antrim are fully accounted for by the human errors described in Sections III and IV of this case study.

13.8 — The Central Lake Marihuana Initiative: The One Uncorrected Item

Halderman’s analysis identifies exactly one contest in Antrim County whose final certified outcome may still be wrong. It is a hyperlocal one.

The Central Lake Village Marihuana Retailer Initiative appeared on the Central Lake ballot. The initial election-day poll tape (November 3) showed 262 Yes / 262 No — a tie, which under Michigan law means the initiative fails. The Central Lake scanner rescanned its ballots on November 6 using the revised election definition. The rescan poll tape showed 262 Yes / 261 No — a one-vote margin, which means the initiative passes. The final certified results reflect the rescan: 262 Yes / 261 No.

Halderman reconstructed the individual ballot-level records from the EMS CastVoteRecord table. The reconstruction shows that three ballots were scanned on November 3 but do not appear in the Nov. 6 rescan. The most likely explanation, per the Central Lake clerk, is that three damaged ballots were rejected during the rescan and the votes were transcribed to new ballots — but those transcribed ballots were apparently not included in the final numbers. It is also possible that the three ballots were simply missed during the rescan due to human error.

Of the three ballots in Halderman’s reconstruction: one is from a precinct not affected by the initiative; one appears blank; the third shows a “No” vote on the marijuana initiative. If that third ballot is valid and was omitted by error, the correct outcome of the contest is a 262-262 tie, meaning the initiative failed.

This is the only item in the Antrim record that Actual Vote would not have caught. And it is worth being precise about why: the error is in the transition from paper ballots to digital tallies on a specific day (the rescan), not in the transition from tallies to reports. It is a counting-layer problem. The poll tape produced by the rescan accurately reflects what the scanner read during the rescan. The scanner during the rescan did not read three ballots that should have been read. No tape-vs.-report comparison can detect a ballot that was never tabulated.

This is the residual uncertainty. It affects approximately one voter’s intent in one hyperlocal contest. It does not affect any race whose outcome has ever been subject to public dispute.

13.9 — What Actual Vote Would Have Contributed

Return to the two clocks from the introduction. The first clock measures how long the wrong number was publicly visible. The second clock measures how long it took for the technical explanation to catch up.

In Antrim, the first clock ran for forty-eight hours. The 4:09 a.m. report on November 4 was the wrong number; the amended unofficial results on November 6 were essentially the final number. During that forty-eight-hour window, the screenshot circulated globally and the narrative established itself.

What AV would have done during that window is simple. AV volunteers photograph the scanner poll tapes at the close of polls. The tapes in Antrim showed the scanner’s own interpretation of the ballots under the election definition loaded on the memory card. For sixteen of the eighteen scanners, that interpretation was under the initial definition — which meant the tapes correctly identified candidates by their alphabetically-assigned names within the initial ID scheme. When the county then posted its 4:09 a.m. report, which reflected the EMS’s re-interpretation of the same memory cards under the revised definition, the two would have disagreed.

AV’s tape-to-report reconciliation would have surfaced three signals within hours:

Direct tape-to-report mismatches at the precinct level. In most of the affected precincts, the tape totals and the county’s reported totals for each candidate would not have matched.
Cross-precinct implausibility. Republican-leaning precincts reporting heavy Biden wins, when AV’s tape-based totals showed those same precincts going heavily for Trump.
Pattern mismatches across races. The ID-shift affected every contest, not just the presidency. AV would have flagged simultaneous anomalies in the U.S. Senate, U.S. House, State Representative, and State Board of Education contests — a pattern that immediately indicates a systemic interpretation error rather than a race-specific one.

Under an AV-equipped response, the forty-eight-hour window during which “Biden won Antrim” could establish itself as an uncontested claim would collapse to the time it took a volunteer to photograph a tape and upload it — a matter of minutes, not days. The screenshot could still have gone viral. But the context — the tape totals, independently captured, showing what the scanners had actually read — would have gone viral alongside it.

This is the function of independent verification. It is not that AV prevents reporting-layer errors. Errors happen; human beings run elections; mistakes will continue. What AV does is prevent the wrong number from enjoying a period of exclusive public visibility. The technical explanation does not need to wait seventeen days for a second-amended certified result, or 142 days for a forensic report, or 231 days for a Senate Oversight Committee finding, or 533 days for a Court of Appeals ruling. The correction — the scanner tape totals, the independent comparison — is available from the moment the polls close.

Beyond the 4 a.m. report itself, AV would also have surfaced the secondary data entry errors that required Antrim’s November 16 and November 21 amended certifications. Michigan law already requires county canvassers to compare certified results to the poll tapes. Antrim’s canvassers did not complete that comparison correctly. The data entry errors affecting 2.6% of county votes persisted into the first certification and were only corrected over the following two weeks. AV provides independent redundancy for the legally-required comparison that Antrim’s canvassers failed to perform.

The Transparency Frame: Catching the Narrative, Not Just the Error

The discussion above is the offensive frame: Antrim is presented as a case where AV would have caught an error that the existing process did not. That description is accurate, but it understates what AV would actually have done in this particular case, because the existing process did catch the error. The county published corrected results within forty-eight hours. The Michigan Department of State’s “Isolated User Error” fact-check appeared three days after the initial report. The state hand count confirmed the corrected results to within twelve votes. The error was caught. By the standard of error-catching alone, the existing apparatus worked — it was simply slow, and the correction did not arrive with evidentiary force equal to the original mistake.

The harm in Antrim was not the original error. The harm was the multi-year narrative that the original error seeded — the lawsuits, the ASOG report, the “Kraken” filings, the materials circulated before January 6, the eventual $787.5 million Fox–Dominion defamation settlement. The harm grew not from the wrong number but from the gap between the wrong number and the technical explanation of what the wrong number actually meant. During that gap, the wrong number had no public counterweight that did not depend on trusting the same officials whose accuracy was being challenged. The Michigan Department of State said the number was wrong. So did the Antrim County Clerk. So did Dominion. Each of those statements was true, and each came from precisely the institutions that the contemporaneous narrative was alleging had produced the wrong number deliberately. The structural problem is the loop: the verifier is the entity whose conduct is in question. A reader inclined to distrust the official record had no independent record to check.

This is the frame in which AV’s contribution to Antrim is most precisely described. AV’s value is not — or not only — that it catches the error. AV’s value is that it produces a record of the precinct-level totals that does not come from the county, the state, the vendor, or any party to the dispute. The poll tape is a physical artifact printed by the scanner at the close of polls. A photograph of that tape, captured by a volunteer at the precinct and uploaded to a public corpus, is a piece of evidence whose independence from the official reporting chain is structural, not asserted. When the 4:09 a.m. screenshot circulated on November 4, the corrected official numbers — which the state would publish on November 5 and 6 — would have arrived alongside (or behind) something more durable: hundreds of citizen-held photographs of poll tapes from across Antrim County, which together establish the correct precinct-level totals as a publicly checkable fact. The reader who distrusts the Secretary of State can still look at a photograph of a tape. The reader who suspects Dominion can still check whether two numbers match.

This is what the case-study methodology calls transparency as an independent value. The accuracy assurance — the catching of the original error — is downstream. The load-bearing benefit is that the evidence on which the accuracy assurance rests is independent of the institutions whose accuracy is being challenged. Antrim is the clearest illustration in the collection of why that distinction matters. The error was caught. The harm came not from the un-catching but from the seventeen-day window during which the wrong number had no independent evidentiary counterweight. AV’s contribution to a future Antrim is not primarily that it would catch the error faster (though it would). It is that the corrected record, when it arrives, would arrive embedded in a citizen-produced corpus of primary-source evidence that is structurally outside the dispute. The narrative can still try to establish itself. What it cannot do, in a world with AV, is establish itself in an evidentiary vacuum.

This re-frame matters for one further reason. The original 4:09 a.m. error in Antrim favored Biden; the corrected results favored Trump. The bad-faith narrative that grew from the screenshot was a narrative pushed primarily by Trump-aligned actors against a Democratic outcome. But the same dynamic operates in the opposite direction in the existing collection: in Prince William County, Virginia, an undetected reporting error favored Trump and shorted Biden by 1,648 votes for more than a year. Had Prince William’s error become a public anomaly during the unsettled days of November 2020 — a Republican county reporting an unusually large Trump margin in a state Biden carried — the same kind of narrative could have grown around it from the opposite direction. The defensive value of independent evidence is not partisan. It is structurally identical regardless of which side of the political spectrum is attempting to convert a numerical anomaly into a story about fraud. AV produces the evidence either way. Whether the evidence ends up confirming or refuting the suspicion is a question for the evidence to answer; what matters is that the evidence exists, independently, before the question is asked.

What AV Would Not Have Done

AV would not have caught the Central Lake Marihuana Retailer Initiative error. That is a counting-layer discrepancy — three ballots missing from the rescan tabulation, not an error in the transition from tabulation to report. An RLA of that specific contest, or a hand count, could surface it; AV cannot.

AV would not have determined whether the 4 a.m. error was deliberate or accidental. That is a forensic question, which Halderman’s 142-day investigation answered. AV’s function is to flag discrepancies; investigation of cause is separate and necessarily slower.

AV would not have answered the broader questions about Dominion’s software architecture — the ChoiceManifestation ID design that made a single write-in-blank addition capable of cascading across the entire county, the lack of a built-in check for election-definition compatibility between scanner cards and EMS, the weak authentication and access controls on the EMS. Those questions require the kind of rigorous peer-reviewed technical research Halderman conducted, and they produced Halderman’s twenty recommendations to Dominion, Michigan, and the broader elections community. Research like that is essential. It is also, necessarily, slow. AV’s role is to provide verification at the speed of the election itself, so that the ballot count can be trusted before every open research question is resolved.

13.10 — The Relationship to Other Audit Methods

Antrim 2020 is also useful for illustrating how Actual Vote fits alongside the verification methods that already exist.

Logic & Accuracy testing (L&A). Every Antrim scanner was L&A tested in October 2020. The testing did not catch the problem. Two of the three affected townships (Central Lake and Warner) produced test poll tapes that reflected the initial ballot designs — which, under the revised definition loaded into the EMS, would produce wrong results. The testers either failed to notice the discrepancies on their test tapes or noticed them and did not flag them up. Michigan’s L&A testing is per-scanner; it does not involve loading the test results into the EMS and verifying the final report. Halderman’s Recommendation #2 is that states require such “end-to-end” L&A testing. AV is not a substitute for L&A; it catches errors that slip past L&A, which is exactly what happened here.

The hand count. The State Bureau of Elections’ December 17, 2020, hand count of the Antrim presidential contest confirmed the certified totals to within twelve votes. This validated that the counting layer — the scanners reading paper ballots — was accurate. The error lived in the reporting layer, which is exactly where AV operates. Hand counts are slow, expensive, and only practical in contested races. They are the strongest form of verification when they happen. AV provides a much cheaper, faster, universally-applicable form of verification that catches errors the counting layer doesn’t have.

Risk-limiting audits (RLAs). Michigan’s Antrim hand count was not an RLA — it was a full hand count of one contest. But the principle applies: an RLA statistically verifies that the scanners counted the paper ballots correctly. It does not verify that the scanner results were faithfully transmitted into the official report. The Central Lake Marihuana Retailer Initiative could have been surfaced by an RLA of that specific contest; it could not have been surfaced by AV. The 4 a.m. EMS error would have been surfaced by AV immediately; it might or might not have been surfaced by an RLA, depending on whether the RLA procedure compared reports to tape totals. RLAs and AV are complementary. Both should exist.

The canvassing process. Michigan already legally requires county canvassers to compare certified results to poll tapes. Antrim’s canvassers failed to do so. One lesson of Antrim is that a legal requirement is not self-enforcing — someone has to actually perform the comparison. AV adds an independent layer of the same comparison, performed by volunteers outside the canvassing apparatus. When the canvassers do their job, AV is redundant. When they don’t, AV is what catches the error.

13.11 — What We Know and What We Don’t

What we know:

On the morning of November 4, 2020, Antrim County published a wildly incorrect unofficial results report. The root cause was a mismatch between election definitions: the scanners in most townships ran under the initial definition while the EMS ran under the revised definition, a mismatch triggered by last-minute ballot design changes in three specific townships and by county staff’s failure to distribute the updated memory cards to all affected scanners. The county corrected the results over the following seventeen days through three amended reports. A state hand count confirmed the final corrected totals to within twelve votes. Every subsequent institutional review — by a GOP-led state Senate committee, by a trial court, by a unanimous Court of Appeals panel that included two Republican-appointed judges, by the state supreme court, by an academic forensic investigation that won best-paper recognition at USENIX — reached the same conclusion: the initial error was human, the corrected results were accurate, and the security and design claims made by the contemporaneous ASOG report were technically wrong on nearly every particular.

What we don’t know:

Whether the three Central Lake Village ballots that Halderman identified as appearing in the November 3 scan but not in the November 6 rescan were omitted due to legitimate invalidation or due to human error. The paper ballots are unavailable for re-examination. If the omission was due to error, the correct outcome of the Central Lake Marihuana Retailer Initiative is a tie, meaning the initiative failed, rather than the 262-261 “passage” currently in the certified results. This is the one item in the Antrim record that remains genuinely uncertain.

Whether any similar ballot-design-mismatch errors occurred elsewhere in 2020 — in Michigan or in other states using Dominion or comparable systems — and were caught through existing processes before they produced visible anomalies. Halderman recommends that other voting-system vendors review their equipment for the same kind of error. As far as we know, no comprehensive such review has been performed publicly.

Whether the L&A testers in Central Lake and Warner noticed the test-tape discrepancies and ignored them, or simply failed to notice. The testing documentation survives but does not capture what the testers saw and judged.

13.12 — The Companion Case: Prince William County

This case study has a mirror. See case study 03, Prince William County, Virginia, 2020.

In Prince William County, the scanners counted correctly and printed correct tapes, but the process of reformatting tape data for the Virginia state reporting system introduced approximately 3,975 votes of error in the presidential race: Biden shorted by 1,648, Trump over-reported by 2,327. The error was not caught by the county’s certification process. It was not caught by the Virginia statewide risk-limiting audit. It sat in the official record for more than a year, until a new registrar noticed an anomaly while responding to a public records request.

In Antrim County, the scanners counted correctly and printed correct tapes, but the Election Management System misinterpreted the memory cards when producing the aggregate report because it ran under a different ballot definition than the scanners. The error was noticed immediately because the initial report was flagrantly impossible. The correction process, however, took seventeen days to finalize.

In Prince William, the error favored Trump. In Antrim, the error favored Biden.

In Prince William, the error persisted silently for a year. In Antrim, the error was corrected within two days but spawned a conspiracy narrative that lasted years.

In Prince William, the statewide RLA missed it because the RLA checked counting, not reporting. In Antrim, the statewide hand count caught it because a full hand count catches both.

In Prince William, the reform was the county publishing its scanner tapes online. In Antrim, the remedy was a state hand count, a Senate investigation, a forensic analysis, and a multi-year litigation process.

Both are reporting-layer errors. Both are exactly what AV is designed to catch. Both would have been flagged immediately by independent tape-to-report reconciliation. The fact that the errors ran in opposite directions is not an artifact of the cases being cherry-picked. It is an artifact of what reporting-layer errors actually are: technical artifacts that run in whichever direction the misinterpretation happens to run.

A verification method that only helps one side is not a verification method. Actual Vote either works in both directions or it doesn’t work at all. Prince William and Antrim together demonstrate that it works in both directions.

13.13 — Why It Matters

Antrim 2020 is the case that illustrates why the speed of verification matters as much as its accuracy.

The technical explanation of what happened in Antrim was available quickly in a narrow sense: the Michigan Department of State published its “Isolated User Error” fact-check on November 7, 2020, three days after the initial report. But “available” is not the same as “established.” The forensic report took 142 days. The GOP-led Senate Oversight Committee’s agreement took 231 days. The unanimous appellate ruling took 533 days. The final state supreme court decision took 765 days. The full public release of Halderman’s report took roughly 900 days.

During those years, the 4:09 a.m. screenshot was an uncontested empirical fact for millions of people. The narrative built on it — that Dominion had stolen the election — shaped a national political movement, motivated dozens of lawsuits, entered materials circulated before January 6, and was the central factual claim at issue in a defamation case that ended with Fox News paying Dominion $787.5 million. This is not an abstract cost of slow verification. It is a measurable cost, payable in institutions, in public trust, and ultimately in dollars.

The underlying failure mode was narrow and regional: three ballot design changes, incomplete distribution of updated memory cards, an EMS that could not detect an election-definition mismatch at load time. The downstream political impact was national and years-long. That asymmetry is the case for independent verification.

Independent verification is the thread that runs through every case study in this set. Prince William County’s reform was to publish its own scanner tapes online, which is the government version of what AV does. Shelby County’s reform (see case study 01) was hand-counting. Monmouth County (case study 02) used RLAs. In each case, the underlying idea is the same: the official record should be checkable against a source independent of the system that produced it. The scanner poll tape is that source. AV makes the poll tape publicly and independently available at the moment of the election, not at the end of a multi-year investigation.

An AV-equipped November 4, 2020, would not have prevented the ballot-definition mismatch. It would not have prevented Antrim from publishing a wrong number at 4:09 a.m. What it would have done is make the correction available on the same clock as the error. The tape totals — independently photographed, independently posted, independently verifiable — would have been available alongside the EMS report. The screenshot could still have gone viral. What could not have happened, in a world with AV, is the forty-eight-hour period during which the wrong number had no public counterweight. And that forty-eight-hour period is where the narrative established itself.

The counting machines worked. The paper ballots were accurate. The hand count confirmed the scanner totals to within 0.1%. What failed was the translation from what the scanners had read to what the public saw. That failure is what Actual Vote exists to catch — not retroactively, not after multi-year litigation, but on the same night as the election itself, at the speed of the people watching it happen.

13.14 — Further Reading