Case Study Methodology

Why we use case studies

America Counts exists to improve election transparency through independent verification of vote reporting. Our Actual Vote system provides that verification. But to explain why independent verification matters — why it’s worth people’s time and effort to use the Actual Vote app, why it’s worth donors’ support, why it’s worth the attention of policymakers and election administrators — we need to show what happens when vote reporting goes wrong and no independent verification exists.

The only way to do that is with real examples. Without them, the claim that “vote reporting is vulnerable to error and fraud” is an abstraction. With them, you can see exactly what goes wrong, what questions get left unanswered, and what independent evidence could have clarified.

That’s what our case studies do. Each one examines a real situation involving a documented vulnerability, error, or accusation of fraud in vote reporting, and asks: what happened, what do we know, what don’t we know, and how would Actual Vote have changed the picture?

Non-partisanship and case studies

You might wonder why a non-partisan organization is featuring stories that involve specific parties, candidates, and elections. It’s a fair question, and one we spent a lot of time on internally.

The short answer is that these stories are the only way to concretely demonstrate the problem Actual Vote exists to solve. The longer answer has four parts.

First, the case for Actual Vote is inherently non-partisan, even though the evidence for it involves partisan elections. America Counts’s invitation to everyone to use Actual Vote is structurally non-partisan — the system produces independent evidence that benefits all parties equally. But to make the case that independent verification is necessary, we must present real evidence of vulnerability: real situations in which vote reporting produced wrong or questionable results, and in which independent evidence would have helped resolve the questions. Individual stories inevitably involve specific parties and candidates. The collection of stories, taken together, establishes something that transcends any individual party: that vote reporting in the United States is vulnerable to error and fraud, and is not sufficiently audited. That conclusion is the foundation for everything Actual Vote does.

Second, it would be irresponsible for experts not to synthesize this information. Vote reporting is a complex technical and administrative process. Evaluating whether a particular incident represents error, fraud, or normal operations requires expertise that most voters don’t have and shouldn’t need to acquire. America Counts has that expertise. Part of our non-partisan mission is to synthesize known information about the state of vote reporting in the United States, present it honestly, and let people form their own informed opinions about whether independent auditing is worthwhile. Refusing to present this evidence — because individual stories happen to involve members of one party or another — would be a disservice to the public and a failure of our mission.

Third, we select case studies for methodological relevance, not partisan impact. Our selection criteria are: the case should clearly illustrate a vulnerability in vote reporting; the case should be well-documented with accessible sources; the case should involve a situation where Actual Vote’s methodology would have been applicable (or where the limits of the methodology are instructive); and the collection should represent a range of circumstances — different states, different decades, different types of vulnerability, and different parties affected. We try to achieve reasonable balance across the collection, recognizing that perfect theoretical balance is impossible given the cases that history has produced and documented.

Fourth, our annotations stick to documented facts. We present what the sources say happened, we cite those sources, and we withhold judgment on any accusation that hasn’t been definitively resolved. The point of each story is never “this party cheated.” It’s always “this is a situation where independent evidence from Actual Vote would have given everyone, regardless of party, a clearer picture of what happened.”

For a deeper discussion of how America Counts maintains its non-partisan commitment, see our page on Non-Partisan Election Advocacy. For why the Actual Vote system itself is structurally resistant to partisan abuse, see Why Actual Vote Is Trustworthy.

Why this matters in the current political moment

The foregoing is a values statement: we are non-partisan because it is right, and because it is the only posture consistent with our mission. That argument has been made on principle, and it stands. But it is increasingly heard — by readers of every political persuasion — as a kind of institutional naïveté: a pose of neutrality that fails to grapple with the reality that every verification authority in American public life has, at this point, been captured by one side or the other in the perception of the other side. It is worth explaining why, in the current political moment specifically, non-partisan independent verification is not just ethically right but functionally the only kind of verification that can actually land with readers across the spectrum.

The anthropologists David Graeber and David Wengrow, in The Dawn of Everything (2021), describe a process they call schismogenesis: the way cultures come to define themselves in opposition to neighboring cultures. “They are the kind of people who have kings, so we are the kind of people who do not.” “They are the kind of people whose leaders never work with their hands, so we are the kind of people whose leaders do.” The defining move is opposition, not evaluation. What the other culture is for, we are against — and vice versa, often to the point that the original content of the positions becomes nearly irrelevant compared to the fact that the opponent holds them. The writer Cory Doctorow, applying this frame in his essay “Schismogenesis” to U.S. politics during and after the Trump years, observed that the same dynamic had flipped long-standing partisan positions on voting machine companies, pharmaceutical institutions, and federal intelligence agencies: as one side came to attack these institutions, the other pivoted to defending them, in each case not because the underlying evidence had changed but because the positions were now defined by who was against them.

This is the condition under which American election verification currently operates. Every existing verification mechanism is compromised not primarily by its actual performance but by whose apparatus it is perceived to belong to. State election boards are dismissed by one side as captured; federal oversight is dismissed by the other. Risk-limiting audits are dismissed by critics who don’t trust the auditors. Hand recounts are dismissed by critics who don’t trust the counters. Forensic analyses are dismissed by critics who don’t trust the forensics. The problem is not that any of these mechanisms is necessarily wrong. The problem is that, in a schismogenetic environment, the identity of the verifier determines whether the verification is believed — and there is no verifier whose identity is trusted by both sides. The gap between “the result has been verified” and “the result has been verified by someone I trust” has, for a growing share of the public, become unbridgeable.

This is the specific gap that Actual Vote is structurally positioned to close. AV produces evidence — photographs of poll tapes captured at the close of polls — that is independent of every verification authority whose identity has been captured by the schismogenetic process. The tape is a physical artifact produced by the scanner at the precinct. It is photographed by volunteers who are not a partisan apparatus. The comparison between tape totals and officially reported totals is arithmetic, not judgment. A reader who distrusts the state election board can still look at a photograph of a poll tape. A reader who distrusts the federal election oversight apparatus can still check whether two numbers match. The evidence does not require trusting an institution. It requires reading a receipt. In this sense, independent tape-based verification is not just ethically non-partisan — it is functionally non-partisan. It sidesteps the schismogenetic trap by providing evidence that does not route through any of the captured institutions.

It is worth being clear that schismogenesis is a diagnosis of a particular political moment, not a permanent feature of democratic life. Actual Vote was not designed only for this moment, and its value does not depend on this moment continuing. If the current polarization eases, AV still provides real-time independent verification of vote reporting, which is a public good in any political climate. What the current moment does is sharpen the case: the gaps that schismogenesis opens in ordinary institutional verification — the gaps where readers who need convincing cannot be convinced by any source their own side calls captured — are exactly the gaps that independent, tape-based, volunteer-captured evidence was designed to fill. The case for Actual Vote is as strong as it has ever been because the moment is as difficult for ordinary verification as it has ever been. A methodology that provides evidence outside the loop is the methodology this moment actually requires.

Our epistemological posture

America Counts’s case studies are written with deliberate intellectual humility. We believe this posture is not a weakness but a strength — it is what distinguishes credible analysis from advocacy, and it is what makes our work trustworthy to readers across the political spectrum.

We go with what the sources say

Every factual claim in our case studies is drawn from identified sources: court records, official investigations, academic research, contemporaneous journalism, government reports, and primary documents. When we describe what happened in a particular case, we are describing what the best available sources say happened. We cite those sources so readers can evaluate them independently.

When sources conflict — as they sometimes do on dates, numbers, or the sequence of events — we note the conflict and either go with the most authoritative source or present both versions and explain the discrepancy. We do not silently choose the version that makes the better story.

We say “I don’t know” when warranted

Not every case has a clear resolution. In some cases, the sources provide a detailed picture of what happened but not why. In others, the evidence points in a direction but doesn’t get there. In still others, the available evidence simply isn’t sufficient to determine whether what occurred was error, fraud, or something else entirely.

When that happens, we say so. We do not fill gaps in the evidence with speculation, inferences presented as facts, or narratives that assume the conclusion. If the evidence doesn’t support a confident statement, we make a qualified statement and explain what would be needed to resolve the uncertainty.

This matters because elections are high stakes. Asserting something we can’t support undermines the credibility of everything else we say. Acknowledging uncertainty — clearly and specifically — builds the kind of credibility that our mission requires.

We withhold judgment on unresolved accusations

Some of the cases in our collection involve accusations of fraud or deliberate manipulation that have never been definitively resolved — not adjudicated in a court of law, not confirmed by an official investigation, not established beyond reasonable dispute. In those cases, we present the accusations, the evidence for and against them, and the fact that they remain unresolved. We do not take sides.

This is not because we think every accusation is equally likely to be true or false. It’s because our role is not to adjudicate disputed elections. Our role is to demonstrate that situations like these arise, that they leave lasting questions, and that independent evidence from Actual Vote would have helped resolve them — regardless of which direction the resolution pointed.

We distinguish between what AV can and can’t catch

Actual Vote is a powerful tool, but it is not omniscient. It verifies vote reporting — the accuracy of the numbers transmitted from precincts to central authorities and the public. It does not verify vote counting — whether ballots were filled out freely, whether the counting process itself was accurate, whether ballot boxes were stuffed. And it does not verify ballot assignment — whether the voter standing at the check-in desk was actually handed the ballot style they were entitled to receive. (The reporting-vs-counting distinction is the “matched poll tape caveat” discussed in detail elsewhere in the manual; the ballot-assignment distinction is described in the three-layer taxonomy below.)

In each case study, we provide an honest assessment of what Actual Vote would have detected and what it would not have detected. When a case involves counting-layer fraud or ballot-assignment-layer error rather than reporting-layer fraud, we say so and explain why AV’s methodology wouldn’t have caught it. When a case involves a type of evidence (such as DRE machines that may not have printed poll tapes) where AV’s applicability is uncertain, we flag the uncertainty.

This honesty serves two purposes. It prevents readers from developing unrealistic expectations about what Actual Vote can do. And it builds trustworthiness: a reader who sees us acknowledge limitations will have more confidence in our claims about what AV can do.

We are deliberately conservative about historical claims

Some of the cases in this collection are historically contested, often sharply. Our voice is deliberately conservative about historical claims, for three reasons. First, conservative claims age well: a claim that is sourced and qualified is robust to subsequent re-examination, while an overreaching claim is fragile. Second, the case for Actual Vote stands on its own — it is supported by recent, well-documented cases and by general structural principles, and does not require any contested historical episode to be characterized one way or another. Third, our credibility is built on rigor, and asserting contested claims as settled would undermine it. Where the evidence is strong (as with the LBJ 1948 case, documented at length by Robert Caro), we say so. Where the evidence is contested (as with Ohio 2004), we say that too.

The three-layer taxonomy: where the error lives matters

One of the most difficult aspects of building our case study collection is that it can be surprisingly hard to determine, from conventional sources, where in the election process a given incident actually occurred. Sometimes you simply can’t tell from the available information. That determination matters because Actual Vote’s methodology operates at a specific layer, and the layer an error lives in tells you whether AV can detect it, whether a recount can repair it, and what kind of safeguards the incident actually points toward.

We find it useful to think of the election process in three layers. The counting-versus-reporting boundary — the poll tape — is developed in detail in the manual’s Conceptual Foundation; what follows adds the ballot-assignment layer and the implications each layer carries for the case studies:

Ballot-assignment layer. The moment a voter is identified, matched to an address, and handed (or shown) the ballot style they are entitled to vote. In a traditional precinct-based polling place, this layer is physically protected by the fact that each polling place stocks only one ballot style. In a vote-center model that uses electronic pollbooks to print ballots on demand, this layer is governed by software, and software can fail. A ballot-assignment failure means that some voters cast ballots in races they were not entitled to vote in, while being deprived of the chance to cast ballots in races they were entitled to vote in. By the time the ballot reaches the scanner, the error is already encoded in the paper. By the time the poll tape prints, it is already baked into the tally. AV cannot detect this layer of failure, because every subsequent layer (counting, reporting) is internally consistent with the (corrupted) paper trail. The clearest documented example is Williamson County, Texas, 2020, where a print-queue race condition in the county’s electronic pollbooks handed some voters ballots encoded for the wrong districts; the scanners read those wrong ballots correctly, every downstream layer was internally consistent, and no recount could recover the error. We cite Williamson here, at the boundary of the taxonomy, rather than as a standalone case study, precisely because it sits outside AV’s reach. AV would have helped with one downstream symptom — the county produced no precinct-level totals for roughly eight months, and independent poll-tape capture would have supplied those on election night — but Antrim County, Michigan, 2020 is a cleaner illustration of AV’s timeliness value (a wrong number caught in real time rather than a reporting void filled), so we use Antrim for that point and reserve Williamson for marking the ballot-assignment boundary.

Counting layer. The transition from a marked paper ballot to a digital tally inside the scanner — miscounts, ballot-box stuffing, misread ballots, configuration errors. AV does not, in general, verify this layer; hand counts and risk-limiting audits are the established methods for doing so. There is, however, a partial exception, illustrated by DeKalb County, Georgia, 2022: when a counting-layer error is severe enough that the poll tape itself becomes identifiable as wrong against ground truth — a candidate showing zero votes in the precinct where she demonstrably voted, a turnout figure manifestly inconsistent with check-in records, a candidate-distribution pattern impossible given known support — the content of the tape, made publicly available by AV, can flag the error even though AV’s standard tape-vs-result comparison would show no discrepancy. This is an additive capability rather than a replacement for hand counts and RLAs, and it is reliable only for counting-layer errors large enough or patterned enough to be visible on inspection. But it expands the practical envelope of what poll-tape transparency catches, and the DeKalb case is the collection’s documentation of the boundary.

Reporting layer. The chain from the poll tape produced at the precinct to the officially published county, state, and national results — dropped data transfers, transposed entries, aggregation and formatting errors, and at the extreme, deliberate alteration of totals. AV is designed specifically to verify this layer: a photograph of a poll tape at the close of polls, captured independently of the government reporting chain, provides the evidence against which reported results can be checked. Prince William County 2020 and Antrim County 2020 are the collection’s clearest illustrations of this layer.

In practice, the available sources for a given incident often don’t make these distinctions clear. There are understandable reasons for this. Journalists covering election controversies are typically writing for a general audience and don’t have reason to specify exactly where in the chain the alleged problem occurred. Official investigations may focus on the outcome (were the reported results correct?) rather than the mechanism (at what step did they become incorrect?). Academic analyses may treat the entire process from ballot to final result as a single system. Legal proceedings focus on whether the election outcome should change, not on the technical taxonomy of where the failure occurred. And election officials themselves may not know — or may not publicly specify — exactly which step in the process malfunctioned.

All of these are reasonable for their respective purposes. But for America Counts, the distinction is critical. A case that involves a reporting-layer failure is a direct illustration of why Actual Vote is needed — it shows the exact type of problem that AV detects. A case that involves a counting-layer or ballot-assignment-layer failure is important for the broader picture of election vulnerability, but it doesn’t directly demonstrate AV’s value, and presenting it as though it does would be misleading. Layer-honesty is what allows the case for AV to be credible. A methodology that claims to catch everything catches nothing in the reader’s mind. A methodology that is clear about which layer it verifies — and which layers require other safeguards — earns the trust it asks for.

This is why we invest significant effort in researching each case study to the point where we can make this determination with reasonable confidence — or, when we can’t, to be transparent about the ambiguity. It’s also why we prioritize cases where the reporting-layer relevance is clear and straightforward. Cases at other layers are included when they illuminate something important about the problem space. Where a failure sits entirely outside AV’s reach — like the Williamson County ballot-assignment failure discussed above — we cite it to mark the boundary rather than presenting it as a standalone case study, so the cases in the collection all speak directly to what AV verifies.

What we’re trying to illustrate

The central purpose of our case studies is to demonstrate, through concrete real-world examples, that vote reporting in the United States is vulnerable to unintentional error and deliberate fraud — and that this vulnerability creates situations where elections lack the evidentiary standard that a healthy democracy requires.

This is not a speculative or theoretical claim. Our case studies span more than a century of American elections, from 19th-century machine politics to 21st-century software bugs, from rural Texas precincts to major metropolitan counties, from both parties and every region of the country. They include:

Equipment failures that silently drop votes during data transfer, producing wrong results with no error indication (Gaston County 1998, Butler County 2008)
Software bugs that exist for a decade without detection or disclosure by the vendor (Butler County 2008, Humboldt County 2008)
Reporting anomalies of consequential magnitude that are detected only by accident (Florida 2000)
Situations where the evidence base is insufficient to determine what actually happened, leaving lasting questions about whether the official results were correct (Baldwin County 2002, Warren County 2004)
Cases where independent evidence — sometimes as simple as a single photograph of a poll tape — uncovered discrepancies that official processes had missed (Shelby County 2015)
Cases where transparent, public recovery from equipment failure produced high confidence in the results — demonstrating the positive outcome that independent verification enables (Gaston County 1998)

Each of these situations illustrates the same fundamental point: when vote reporting lacks independent verification, errors may go undetected, accusations may go unresolved, and the public’s ability to rationally evaluate whether the official results are accurate is compromised.

The evidentiary standard problem

A recurring theme across our case studies — and a central motivation for Actual Vote — is what we call the evidentiary standard problem: elections are among the highest-stakes collective decisions a society makes, but the evidentiary basis for believing the results are accurate is often thinner than those stakes demand.

What should that standard be? At its core, a healthy democracy requires that any reasonable person — regardless of party, ideology, or which candidate they supported — should be able to look at the available evidence and rationally conclude that the reported results accurately reflect the votes that were cast. This doesn’t mean the evidence must eliminate all conceivable doubt; it means the evidence must be strong enough, transparent enough, and independently verifiable enough that the conclusion “the results are accurate” is the natural, well-supported inference rather than an act of faith or institutional trust.

In most elections, in most jurisdictions, this standard is probably met — the results are accurate, the process worked, and the official totals reflect reality. But “probably” is doing a lot of work in that sentence, and “most” is not “all.” The problem isn’t that every election falls short. It’s that when an election does fall short — when the evidence is insufficient, when questions arise that can’t be answered, when the process is opaque at a critical step — there is often no way to resolve the situation because the independent evidence that would resolve it was never created.

And crucially, the evidentiary standard matters even when the winner of the race is not in question. A race that was won by 20 points can still have a 4,000-vote reporting discrepancy in a single county. That discrepancy doesn’t change who won, but it matters: it means the official results don’t accurately reflect the votes that were cast. It means the margin of victory is wrong. It means that voters in the affected precincts had their votes miscounted in the public record. And it means that whatever caused the discrepancy — equipment failure, human error, or something else — went undetected by the existing safeguards and could cause a more consequential problem in a closer race. An evidentiary standard that only cares about discrepancies large enough to change an outcome is not a standard at all; it’s an acceptance that some amount of inaccuracy is tolerable as long as it doesn’t happen to matter this time.

In many of our cases, the problem is not that the election was definitely wrong. It’s that we don’t have enough evidence to be confident it was right. When a county reports zero votes from a third of its precincts due to a data cartridge failure, and the recovery depends on one official’s decision to open the process to the public — what happens in the county where the official doesn’t make that decision? When a software bug silently drops votes for a decade across 34 states — how many elections were affected that no one ever checked? When 80,000 votes are alleged to have been switched in a county where observers were locked out — and no independent evidence exists to confirm or refute the allegation — what is the public supposed to believe?

In each of these situations, the absence of independent evidence is itself the problem. Not because the election was necessarily wrong, but because the public has no rational basis for being confident it was right. This gap — between the evidentiary standard that high-stakes elections deserve and the evidentiary standard that currently exists — is what Actual Vote exists to close.

Actual Vote doesn’t just help catch errors and fraud after they occur. It also helps in the more common situation where vote reporting was done correctly: in that case, the independent evidence from Actual Vote constitutes confirmation that the results are accurate. This is valuable in its own right. In an era of declining public trust in institutions, having independently verifiable evidence that the reported results match the precinct-level counts is a public good that benefits everyone — voters, candidates, administrators, and the democratic process itself.

Transparency as an independent value

Most of our case studies are framed around the question of whether an election was counted and reported accurately. That question is important, and it is the question that most naturally drives interest in independent election evidence. But it is not the only question, and framing Actual Vote purely as an accuracy-assurance system understates what it actually does. Actual Vote is, first and most fundamentally, a transparency system. Accuracy assurance is one of several things that independent transparency enables; it is not the only one, and in some cases it is not even the most important one.

There are two reasons this matters. The first is defensive. In a contested political environment, correct official results are routinely subjected to bad-faith accusations of error or fraud. When that happens, the official record is not always sufficient to rebut the accusation — not because the record is wrong, but because the record is produced entirely by the same entities whose accuracy is being challenged. The accusation can always be met with a new version of itself: of course the officials say the totals are correct; the officials produced the totals. Independent evidence — photographs of the precinct-level poll tapes, taken by ordinary citizens on election night, before any aggregation has occurred — breaks that loop. It provides a source of truth that is not produced by election administrators, not produced by the political parties, and not produced by the candidates. When the official reporting is correct, independent evidence confirms it is correct, and does so in a way that a person inclined to distrust the official record has a harder time dismissing. A case in which Actual Vote evidence helps establish in court that the official reporting was accurate would be, from a transparency standpoint, just as powerful as a case in which Actual Vote evidence helps establish that the official reporting was wrong. The valence is different — one outcome is vindicating, the other is disruptive — but the underlying public good is the same: independent evidence, available to all, sufficient to rationally resolve the question.

The second reason is deterrent. The mere existence of a broad, citizen-produced corpus of precinct-level tabulator tape photographs changes the incentive structure for both actual bad actors and bad-faith accusers. If you are contemplating manipulating vote reporting at the precinct level, you now face the possibility that the primary-source evidence of the correct count has already been photographed and preserved by dozens of independent observers, and that your manipulation will be detectable by a cross-reference. If you are contemplating making an unfounded accusation of fraud in order to cast doubt on a result you don’t like, you now face the possibility that anyone can cite the independent evidence — or perform their own analysis on it — to demonstrate that the official reporting matches the precinct tapes. Neither of these effects depends on Actual Vote catching any specific problem. They depend only on the existence of the corpus and on its being known to exist.

Our North Carolina 2024 analysis is the archetypal transparency case in this sense. North Carolina 2024 was, as far as anyone can tell, an ordinary election. No specific allegations of error or fraud attached to it. But all elections deserve independent audits, and we happened to obtain Actual Vote videos from precincts in that election. We did the analysis — a straightforward cross-reference of the tabulator-tape totals against the county- and state-level official reporting — and discovered that they matched exactly. We published the result. Accuracy was never in question; nothing was wrong; there was nothing to catch. But the transparency benefit is real and worth naming. Had any bad actor been tempted to call the North Carolina 2024 results into question for bad-faith reasons after the fact, the existence of the Actual Vote evidence — and our published cross-reference — would have functioned as a preemptive check. Anyone could have cited the report, or re-run the analysis themselves, and produced independent evidence that the official reporting was indeed correct. The accusation would have had to grapple with evidence not produced by election administrators, and that is a meaningfully harder thing to do than to grapple with a denial from the officials whose conduct is being challenged.

This is why we think of every case study in this collection as a transparency case — including the ones that are primarily about accuracy. Whenever Actual Vote evidence is in play, transparency is what is being delivered, because the evidence is independent of the official record. Sometimes that independent evidence will confirm the official record, sometimes it will contradict it, and sometimes the question of whether it confirms or contradicts will itself be the disputed point. In all three situations, the public is better off than it would be if the only evidence of what happened were the evidence produced by the same institutions whose performance is in question. Accuracy assurance is a downstream benefit of transparency; transparency is the load-bearing value.

The defensive half of the transparency argument — the use of independent evidence to refute, rather than to expose, claims about election integrity — is taken up at length in Defending Against Bad-Faith Accusations. That file extends the defensive frame across cases the collection already contains (Antrim 2020, Prince William 2020, Florida 2000, Baldwin 2002, Warren 2004, Cook 2016, NC Supreme Court 2020) and adds shorter vignettes drawn from elections in which bad-faith vote-reporting accusations grew in evidentiary vacuums and persisted for years afterward. The vignettes are deliberately bidirectional: bad-faith accusations of vote-reporting fraud have come from both directions of the political spectrum in modern American history, and the value of independent precinct-level evidence is structurally identical regardless of which direction the accusation runs. The methodology principles described in this section are the foundation on which that file is built.

A note on election administrators

Nothing in our case studies should be read as a general indictment of election administrators. The vast majority of election officials at every level — county boards, precinct workers, state election offices — are dedicated, hardworking professionals who take their responsibilities seriously and carry them out with integrity, often under difficult conditions and with inadequate resources.

Our case studies document situations where things went wrong. Some of those situations involved equipment failures that no administrator could have prevented. Some involved software bugs that vendors concealed from the officials using their systems. Some involved the actions of specific individuals who violated the trust placed in them. And some involved situations where the evidence simply isn’t sufficient to determine what happened.

In several of our cases, election administrators are the heroes of the story. Betty McGary in Butler County refused to accept a vendor’s false explanation and pushed an investigation that uncovered a decade-old software bug. The Gaston County board chair opened the recovery process to full public observation, producing confidence in the corrected results. Bennie Smith in Shelby County used a single poll tape photograph to uncover a 40% vote discrepancy.

Actual Vote is designed to complement the work of election administrators, not to replace or undermine it. Independent verification serves the same function as an independent audit in any high-stakes system: it provides an additional layer of confidence that the system is working correctly. When the system is working correctly — which it usually is — independent verification confirms that fact. When something goes wrong, independent verification helps identify the problem quickly so it can be corrected. Both outcomes serve the public interest and the interests of the administrators who run our elections.

Case study structure

Each case study follows a consistent structure:

What happened. A factual account of the incident, drawn from the best available sources. We identify our sources and flag uncertainty where the evidence base is contested or incomplete.

What’s documented. The evidence base: court records, official investigations, contemporaneous journalism, academic research. The strength of the documentation varies by case; we are explicit about which cases rest on strong evidence and which rest on contested or partial evidence.

What’s contested. Competing interpretations, unresolved questions, and areas of genuine disagreement. We do not take sides on contested questions where the evidence is ambiguous.

What Actual Vote would have caught (or missed). An honest assessment of whether AV’s methodology, had it been in use, would have detected the problem described in the case. This assessment accounts for the specific type of problem (counting error, reporting error, or ballot-assignment error — see the three-layer taxonomy above — and silent failure vs. self-announcing failure), the access conditions (were poll tapes posted, were they accessible), and the limitations of AV’s methodology. When AV would not have caught something, we say so.

Why it matters. The connection between the specific case and the broader case for independent vote reporting verification. Not every case proves the same point; some illustrate the need, some illustrate what’s possible, some illustrate limitations.

Selection criteria

We select case studies based on the following criteria, applied to the collection as a whole rather than to any individual case:

Methodological relevance. The case should illustrate something specific about vote reporting vulnerability and/or about how Actual Vote’s methodology applies (or doesn’t apply) to a real situation.
Documentation quality. The case should be supported by credible, accessible sources. We prefer cases with primary documentation (court records, official reports, published investigations) over cases that rely solely on contemporaneous press coverage, though press coverage can supplement primary sources.
Range of circumstances. The collection should represent a range of states, time periods, types of vulnerability (equipment failure, human error, alleged fraud, unexplained anomaly), and parties affected. We aim for reasonable balance, recognizing that perfect balance across every dimension is impossible given the historical record.
Explanatory clarity. The case should be explainable without requiring specialized technical knowledge. Complex cases are acceptable if the core point can be communicated clearly.
Honesty about limitations. We do not exclude cases that present complications or that reveal the limits of AV’s methodology. Cases where AV wouldn’t have helped — or where its applicability is uncertain — are included when they illuminate something important about the problem space.

Our case studies are the evidence for the claim that independent vote reporting verification is necessary and valuable. We believe that evidence is most persuasive when it is most honest.

Current case study collection

As of this writing, the collection includes eighteen case studies spanning more than a century and a half of American elections. Each has been researched from primary sources, written with the epistemological posture described above, and reviewed for factual accuracy and AV relevance.

One useful way to read the collection is by era and type. Historical cases (Tammany Hall, LBJ 1948, Florida 2000, Gaston County 1998) predate or sit at the dawn of the electronic voting era; they illustrate the long history of vote manipulation and are included for context, not as claims about what AV could have done in a pre-digital world. Contemporary pre-AV cases (roughly 2000–2023, from Baldwin County through Monmouth and DeKalb) document vote-reporting problems before AV existed or was deployed; these are the most directly relevant to the case for independent verification, because they show the category of real-world problem AV is designed to catch. Systemic and theoretical cases (Fraction Magic) describe categories of vulnerability rather than specific incidents. AC’s own deployments — the AV-era record — are documented separately on our analyses.

The collection is:

Tammany Hall (1860s–1930s). Machine-era vote manipulation in New York City — ballot stuffing, repeat voting, precinct-level fraud. Primarily a counting-layer case, but illustrates the historical depth of vote reporting vulnerability and the structural incentives that make independent verification necessary.

LBJ and the 1948 Texas Senate Race. The “Box 13” scandal in Jim Wells County — 202 votes added to the tally days after the election, in alphabetical order, in the same ink. A case where the fraud was so crudely executed that even minimal independent verification would have caught it, but no such verification existed.

Gaston County, North Carolina, 1998. Faulty data cartridges caused a third of precincts to report zero votes. The county board opened recovery to full public observation, and results were resolved without protest. The rare positive outcome story — a demonstration of what transparency and public access can achieve.

Florida Presidential Election, 2000. The Volusia County anomaly (negative 16,022 votes for Gore at a precinct with 585 voters), reporting-chain failures across multiple counties, and the fundamental problem of a 537-vote margin in a system with no independent precinct-level verification. The case that most dramatically illustrates the “invisible spread” problem: errors distributed across many precincts are undetectable without precinct-level checks.

Baldwin County, Alabama, 2002. Approximately 6,300 votes vanished from one candidate’s column at a single precinct after election officials reopened the courthouse with the opposing party’s observer gone. Recount was blocked. Ballots were later destroyed by law. The central question — what actually happened — can never be answered.

Warren County, Ohio, 2004. County officials locked down the administration building and barred all observers on election night, citing a terrorist threat the FBI said it had no information about. Part of a broader pattern across Ohio in which 56 of 88 counties destroyed their election records despite a federal retention law and court order. The evidence destruction makes the central questions permanently unresolvable.

Butler County, Ohio, 2008. Betty McGary discovered that a GEMS software bug silently dropped votes during data transfer — the same bug class later found in Shelby County, Tennessee. The vendor blamed antivirus software; independent investigation proved the bug was in the GEMS upload process. Official post-election audits had detected nothing. The case established that a software defect could persist across a decade and 34 states without vendor disclosure.

Shelby County, Tennessee, 2015. Bennie Smith photographed a poll tape showing 546 votes; official results showed 330 for the same precinct — a 40% discrepancy. His single photograph exposed a systemic problem affecting multiple precincts, all in areas with large concentrations of Black voters. The county’s Administrator of Elections retired under pressure. This is the closest real-world demonstration of the Actual Vote concept: one person, one photo, one poll tape.

Cook County, Illinois, 2016. Audit judges tasked with verifying electronic vote counts failed to follow mandated hand-tally procedures, stamped results as verified without performing the required comparison. Election monitoring observers caught and documented the failures. A case about the vulnerability of the human verification layer — even well-designed audit procedures fail when the people executing them don’t follow them.

Prince William County, Virginia, 2020. In Prince William County, Virginia (2020), a ~4,000-vote reporting error survived certification and a statewide risk-limiting audit that confirmed the counting layer with 99%+ confidence — because the error was in the reporting layer, which that audit did not check. It was discovered over a year later, by accident. The textbook reporting-layer failure case.

Antrim County, Michigan, 2020. A mismatch between the ballot definition loaded on scanner memory cards and the one running in the county’s Election Management System caused the 4:09 a.m. unofficial report to swap Trump’s and Biden’s totals in a rural Republican county. The paper ballots and scanner tapes were correct throughout; the error lived entirely at the reporting layer and was corrected over the following days. Between the wrong number and the correction, a viral screenshot established “Biden won Antrim” as a stolen-election narrative that persisted through dozens of lawsuits, a $787.5 million Fox News / Dominion defamation settlement, and years of conspiracy theorizing. The case that illustrates why the speed of independent verification matters as much as its accuracy. Reporting-layer failure; mirror image of Prince William — same failure class, opposite partisan direction.

Georgia Presidential Election (Biden), 2020. Memory card upload failures across four counties left nearly 5,800 votes missing from official results. Every error was a reporting-layer failure — scanners tabulated correctly, tapes printed correctly, but electronic transmission broke down. Caught only because Georgia conducted a full hand recount triggered by the margin. Also includes the Fulton County unsigned tapes discovery (2025): more than 130 tabulator tapes covering 315,000+ ballots were unsigned, and 10 tapes are missing entirely.

North Carolina Supreme Court, 2020. The closest statewide election in NC history: 401 votes out of 5.4 million (0.008%). Four separate counts (canvass, recanvass, machine recount, hand-to-eye sample) produced four different margins — 406, 416, 401, 412 — all confirming the same winner. Neither candidate alleged fraud. The case where the system worked but the question remains whether the checks were complete: none of the four counts systematically verified the reporting chain from precinct tapes to official results across all 2,662 precincts.

Monmouth County, New Jersey, 2022. An ES&S employee’s software reinstallation removed a double-counting safeguard. Six flash drives were loaded twice. The wrong candidate was certified and took office. The official audit reported “tabulation accuracy was 100%.” The error was discovered two months later by a citizen who noticed phantom votes and was initially told he was comparing “apples to oranges.” The only case in the collection where a reporting error definitively changed an election outcome.

Gillespie County, Texas, 2024. After ten of eleven election workers resigned in 2023, the new elections office conducted the March 5 Republican primary largely by hand count. The counting layer changed, in the direction many skeptics of machine counting had long recommended. The reporting layer did not. Workers produced reconciliation errors, duplicated precinct entries, and discrepancies between poll lists and scanner totals that took weeks to resolve. A hand-count election is still a multi-step chain from ballot to publication, and every step from tally sheet onward remains a reporting-layer surface. The case that preempts the “hand counting solves it” argument by showing that the reporting layer is an independent problem regardless of how the counting layer operates.

The Fraction Magic Attack. A demonstrated theoretical vulnerability in the GEMS reporting system: votes stored internally as fractions that round to whole numbers, allowing systematic shaving of one candidate’s totals. Included as an illustration of a sophisticated reporting-layer attack that would be immediately visible through poll tape comparison.

Clarkson v. Lehman (Kansas, 2014–2018). A chief statistician at Wichita State University’s National Institute for Aviation Research identified a statistical pattern in Kansas returns that she believed warranted investigation, and filed an open-records request for real-time audit log tapes from a small number of Sedgwick County precincts. Her request was denied on the grounds that the tapes were “ballots” under Kansas’s criminal ballot-disclosure statute. Her appeal was dismissed as moot in 2018 because the tapes had been destroyed during the pendency of the litigation. The case establishes the access-to-evidence failure mode in its clearest form: a request for primary-source reporting-layer evidence, denied; the denial upheld; the evidence destroyed before merits could be reached. The collection’s pivot from cases of accuracy error to cases of transparency denial, and the most direct illustration of the access regime that Actual Vote is designed to operate outside of.

DeKalb County, Georgia, 2022. A four-step cascade — a mid-cycle redistricting fix, a candidate withdrawal, a ballot marking device text fix, and a procedural mistake in which the third update was built from the wrong base copy — produced a positional mismatch between BMDs (running a three-candidate configuration) and scanners (still expecting four candidates) at 32 of 40 District 2 precincts. Every Election Day vote for Michelle Long Spears was recorded as a vote for Marshall Orson; every vote for Orson was recorded as a vote for the withdrawn candidate, then converted to blank on upload. Spears appeared to have zero Election Day votes in the precinct where she and her husband had voted. She drove to multiple precincts that night photographing the publicly posted poll tapes by hand. A three-day hand recount over Memorial Day weekend produced a reordering of the runoff lineup; Spears went on to win the runoff and take office. The collection’s most distinctive case: the error was strictly speaking a counting-layer failure (the scanners produced wrong poll tapes, and a naive AV poll-tape-vs-official-result comparison would have found them matching), but the error was severe enough that the poll tape itself was identifiable as wrong against ground truth. AV’s poll-tape transparency catches it not by the standard comparison but by making the impossible tape numbers visible to anyone who knows the candidate had at least two votes in her own precinct. The textbook “For Candidates” case in the collection: Spears performed by hand, on a deadline, exactly the workflow AV is designed to systematize.

Patterns across the collection

Several patterns emerge from the collection as a whole:

Individual persistence as the discovery mechanism. In a striking number of cases, errors were caught not by systematic processes but by individuals who noticed something wrong and refused to let it go: Betty McGary in Butler County, Bennie Smith in Shelby County, James Bean in Monmouth County, Eric Olsen in Prince William County, Michelle Long Spears in DeKalb County. Each faced institutional resistance — vendor deflection, official dismissal, bureaucratic inertia, or simply the time-consuming work of assembling enough evidence to be taken seriously — before their concerns were validated. This pattern underscores both the fragility of relying on individual heroism and the potential of systematizing what these individuals did through a tool like Actual Vote. The DeKalb case is particularly striking because the individual in question was the affected candidate herself: Spears performed manually, by driving from precinct to precinct on the evening of and morning after the election, exactly the precinct-level poll-tape comparison that AV is designed to systematize.

Failed audits. Official post-election audits repeatedly fail to catch the errors documented in these cases. Butler County’s audit missed the GEMS bug. Virginia’s risk-limiting audit (99%+ confidence) missed the 4,000-vote discrepancy. Monmouth County’s audit reported “100% tabulation accuracy” while the wrong candidate was in office. Cook County’s audit judges signed off without performing the required hand tally. The pattern is not that audits are worthless — it’s that they test specific things and miss others, and reporting-layer errors fall in the gap.

Evidence destruction. In multiple cases, the evidence that would resolve open questions was destroyed: ballots in Baldwin County (destroyed by law), election records in 56 Ohio counties (destroyed despite a federal court order), Fulton County tapes (missing). AV’s core structural advantage is that its evidence exists outside government custody and cannot be destroyed by the same officials whose work it verifies.

Recurring software bugs. The same class of GEMS upload bug appeared in Gaston County (1998), Butler County (2008), and Shelby County (2015). The DuPage County incident (2004) appears to involve the same failure pattern. Fixes applied in one jurisdiction did not propagate to others. Corporate acquisitions transferred the software through multiple owners without eliminating the underlying vulnerabilities. This pattern suggests that reporting-layer bugs are not isolated incidents but systemic features of a fragmented election technology ecosystem.

The confidence gap. Most of these cases are not about elections that were definitely wrong. They are about elections where we lack sufficient evidence to be confident they were right. The NC Supreme Court case is the clearest example: the system worked, the winner is not in doubt, but the four-count sequence and the unchecked reporting chain leave a gap between “we got the same winner” and “we are fully confident in the margin.” AV’s contribution is not only catching errors — it is providing the evidence base for justified confidence when no errors exist.

Errors run in both directions. Reporting-layer failures are not partisan phenomena. They are technical artifacts that favor whichever candidate the arithmetic happens to favor. In Prince William County, Virginia, 2020, the reporting error overstated Trump’s total by roughly 2,327 votes and understated Biden’s by 1,648 — a net swing of nearly 4,000 votes in Trump’s favor, caught only more than a year later by a new registrar. In Antrim County, Michigan, 2020, the reporting error produced a 4:09 a.m. unofficial total showing Biden beating Trump by 3,260 votes in a rural Republican county — a figure that, when corrected over the following days, reversed to a Trump lead of 3,788. Same failure class. Opposite direction. Both are exactly the kind of error AV is designed to catch. The point is not that these two cases happen to balance each other; the point is that the symmetry is structural. A verification method whose value depends on which party’s candidate is harmed is not a verification method. AV either works for everyone or it doesn’t work at all — and the cases in this collection, taken together, demonstrate that it works for everyone.

Bad-faith accusations also run in both directions. Independent evidence is not only a remedy for errors; it is also a remedy for false claims of error. Sustained bad-faith accusations of vote-reporting fraud are recurring features of American political life, and they have been advanced from both directions of the political spectrum across the cases the collection contains. The post-2000 left-aligned argument that Bush stole Florida and the 2002 left-aligned argument that Riley/Rove stole the Alabama governorship from Siegelman; the post-2004 left-aligned argument that Bush stole Ohio; the 2016 intra-Democratic Sanders-camp argument that Clinton stole the primary; the 2020 right-aligned argument that Biden stole the presidency through reporting-layer manipulation in Antrim, Fulton, Maricopa, Wayne, and other counties — these are different claims, by different actors, against different outcomes, with different mixtures of rigorous and bad-faith analytical components. They share a structural feature: each grew most readily where independent precinct-level evidence did not exist, and each persisted longest where the absence of such evidence made the underlying question unanswerable. AV’s defensive value is bidirectional in the same way and for the same reason its offensive value is. See Defending Against Bad-Faith Accusations for the worked examples.

A note on ongoing development

This collection is not static. As new cases come to light and additional research becomes available, we update existing case studies and consider new additions. Cases under consideration for future inclusion include incidents involving mail-in/absentee ballot reporting vulnerabilities, additional access-to-evidence litigation in states beyond Kansas, and contemporary cases as they emerge. The collection’s goal is not comprehensiveness but representativeness: enough real examples, spanning enough circumstances, to make the case for independent verification concrete and undeniable.