Start Here
Donate

Case 18 — Gillespie County, 2024

18.1 — Introduction

On the morning of Thursday, March 14, 2024, nine days after Gillespie County Republicans had declared their hand-counted primary a success, the county’s elections staff stood around four tables at the county elections office in Fredericksburg, Texas. Spread across the tables were printouts of precinct return totals from all thirteen of the county’s Republican precincts. The staff took turns reading numbers aloud while another worker added them on a calculator. Numbers that had already been summed once on tally sheets, and summed a second time on reconciliation forms, were being summed a third time — this time re-aggregated into the three-column breakdown (Election Day, early voting, mail) that Texas’s state reporting system required. The staff worked through this for nearly three hours. They kept finding new errors, introduced in the very act of correcting the old ones.

The scene sits at the center of this case study because it is a limit case. The Republican Party in Gillespie County had, at considerable expense and in the face of professional advice, chosen to replace electronic tabulation with a hand count — motivated by the conviction that removing machines from the process would produce more trustworthy results. By the morning of March 14, the party’s chairman had already found errors on the reporting forms of twelve of the thirteen precincts. He had already held a canvass at which each precinct judge called out corrected totals. He had certified the results. And still the errors kept coming. An hour after certification, he had to call a volunteer who had left for home — thirty minutes away — to come back and correct one more discrepancy.

The Gillespie case matters for Actual Vote’s methodology in one specific way. The errors it demonstrates were not errors of counting. The chairman of the Gillespie County Republican Party, Bruce Campbell — who had no political incentive to blame his own process — stated publicly that the tally sheets were correct. The errors happened downstream, in the transcription of tally-sheet totals onto reconciliation forms, and in the re-aggregation of reconciliation forms into the state reporting format. They happened in the reporting layer. They were precisely the kind of errors AV is designed to make visible, and they would have been visible to AV independent of whether the counts below were produced by scanners or by people.

The case is in this collection because it preempts a particular argument against AV’s relevance: that the transcription problems AV addresses are artifacts of machine-based counting, and that a return to hand counting would make AV unnecessary. Gillespie is the counter-example. A hand-counted election, run by people who believed deeply that hand counting would eliminate error, produced exactly the kind of reporting-layer errors AV was built to catch. The counting layer changed. The reporting layer did not.

18.2 — Background: The County and the Process

Gillespie County sits in the Texas Hill Country about seventy-five miles west of Austin, with a population of roughly 27,000. Its county seat is Fredericksburg, a tourist town known for its German heritage, its wineries along the Pedernales River, and its location near President Lyndon Johnson’s boyhood home and the state park that bears his name. The county votes heavily Republican. In 2020, the Republican primary there was conducted with forty-five workers, and the results were reported within a few hours of poll close.

In the summer of 2023, the executive committee of the county Republican Party — its membership overlapping substantially with the Fredericksburg Tea Party — voted to abandon the county’s electronic tabulation equipment for the 2024 primary. The decision followed a statewide pressure campaign by election-conspiracy activists who had spent the years after 2020 arguing that machine tabulation was being manipulated by local officials. Most Texas counties that considered the argument declined to act on it, often after analyzing the logistical and financial costs. Gillespie was the only sizable Texas county whose Republican Party chose a full hand count of its 2024 primary ballots. (Travis County Republicans, in neighboring Austin, chose a limited hand count of mail-in ballots only — just under 2,000 of the hundreds of thousands of ballots cast there.)

Under Texas law, primary elections are administered by the political parties themselves rather than by county elections offices. Gillespie’s county elections administrator, Jim Riley, was in the unusual position of providing space, logistical support, and state reporting assistance while having no authority over how the count itself was conducted. When party chairman Bruce Campbell — the man responsible by statute for attesting to the accuracy of the results — made decisions about procedure, staffing, and reconciliation, Riley was a bystander with a professional opinion he was largely not asked to give.

The physical setup for the 2024 primary was divided between a central location for early and mail-in ballots, and thirteen precinct polling sites for Election Day ballots. The central location was The Edge, a tasting room at The Resort at Fredericksburg, a winery complex overlooking a bend in the Pedernales. The tasting room had been a late substitute: the party had originally planned to use a local church, but practice sessions revealed that the church’s acoustics made it impossible for multiple counting teams to work simultaneously without interfering with each other. The winery’s high ceilings and separated seating areas, by contrast, were well suited to the task. Poll workers were paid $12 an hour, as Texas law requires, with the state reimbursing the party and ultimately the county for most of the cost. The thirteen Election Day precinct locations included a Girl Scout cabin, a volunteer fire station, a Farm Bureau office, and an auction house.

The counting methodology within each team was straightforward. Each five-person team had one caller who read out the vote on each ballot, a watcher who confirmed the caller had read the ballot correctly, and three tallyers who kept independent running totals. If the three tallies did not agree, the batch would be recounted. Ballots were handled in batches of approximately fifty, and each team was expected to count one batch in about ninety minutes. The party’s advocates had emphasized the procedure as self-correcting: with three independent tallies per batch, any disagreement would be caught in-team before the count went forward. In principle, the counts themselves were the most rigorously checked step in the process.

18.3 — The Count

Counting began at 7:30 a.m. on Tuesday, March 5, 2024, and continued without a break — Texas law requires the count to be continuous once it begins — for most of the next twenty-four hours. The scale was larger than the 2020 primary by nearly every measure. Turnout was higher than expected, driven by a contested U.S. Senate primary and several hotly fought local races. By the end of Election Day, 8,266 voters had cast Republican primary ballots in Gillespie County, roughly half of them during the early voting period and half on Election Day itself. Each ballot contained more than thirty contests. The total number of paid workers approached 350 across shifts, about seven times the 2020 number.

Campbell had predicted at the start of the night that precinct returns would begin arriving at the county elections office by 8:30 p.m. By 9:30 p.m., none had arrived. Campbell informed Riley that it might be hours before workers finished. Riley’s reported reaction — “Are you kidding me?” — became the headline of the Votebeat story that ran the next morning.

The ballots were not finished being counted until 4:30 a.m. Wednesday. The final precinct to report was Precinct 4, staffed by a first-time election judge named Joy Smith at a Girl Scout cabin — the last of the thirteen precincts to deliver its paperwork. Of Texas’s 254 counties, Gillespie was second-to-last to report its results. At about 5 a.m., Riley, alone in the elections office after the party workers had left, told the Votebeat reporter still watching from the hallway: “You saw how this went. This was a circus.” He added that he would withhold judgment on whether the count itself was accurate, because he had not had eyes and ears in the rooms where the counting happened. Texas law permitted only the counters, poll watchers, and party officials inside; credentialed journalists were excluded.

On the morning of March 6, Campbell and the advocates who had organized the hand count declared the effort a success. A local conservative publication praised the night’s work: hundreds of volunteers, thousands of paper ballots counted within twenty-four hours. The Votebeat coverage that went out that morning noted several concerning observations — the absence of publicly visible quality-control steps, the party’s inability to say how many counters had worked at each precinct or what training they had completed, the speed of the count relative to documented studies of hand-count accuracy — but the results had been returned to the state, the candidates were informed of the outcomes, and the story appeared closed.

18.4 — The Discovery

Scott Netherland had been a Republican election judge in Gillespie County for more than a decade. He had opposed the hand-count plan — believing the existing machines worked — but he had taken his responsibility seriously, running the Precinct 6 count as a professional exercise. By 11 p.m. on election night he had completed his precinct’s count, submitted his reconciliation paperwork to the county elections office, and gone home. He had 197 voters at his precinct on Election Day.

The next morning, uneasy, he pulled out his copy of the paperwork and started checking his math. For each race on the ballot, the votes plus the undervotes should have summed to 197. In one race, he had reported 160 votes. In another, 157. As he went down the list, he hit a third race with 207 votes reported — more than the number of voters at his precinct. In the end, he had misreported the totals for seven separate races on his precinct’s reconciliation form.

“My heart sank,” he told Votebeat.

Netherland called Campbell and drove to the county elections office to review his tally sheets. When he arrived, he began comparing the tally sheets to the reconciliation forms from other precincts that had already been turned in. He realized the errors extended beyond his own. Multiple other precincts had reported obviously incoherent numbers. If the numbers on the reconciliation forms were to be believed, the errors had been introduced somewhere between the tally sheets — which Netherland was satisfied were correct for his own precinct — and the summary forms that had been submitted as the official record of the count.

Campbell, alerted to the problem, spent the weekend before the canvass doing exactly the kind of reconciliation work Netherland had done on his own precinct, but now across all thirteen. He built a spreadsheet. Each row was a race; each column a precinct. Each cell was a reported vote total. At the bottom of each precinct’s column, the sum of votes per race should have reconciled to the precinct’s voter count. Campbell worked down the spreadsheet looking for cells that did not reconcile.

He found them in the reporting forms of twelve of the county’s thirteen precincts. Precincts 2 and 6 — Netherland’s precinct — had errors in seven races each. Precinct 12 had errors in six. Only Precinct 9, under chair Betty Hahn, reported a clean set of numbers; Hahn, who had opposed the hand count in 2024 and would oppose it again in 2026, told a Votebeat reporter about her standards simply: “If I’m going to do something, I do it right.” Campbell highlighted the errors in red, distributed the spreadsheet to the precinct judges, and told them they would need to present corrected numbers at the canvass.

Netherland had not been asked to double-check his work. No party protocol required it. It was not a formal audit step, just a career election worker reviewing his own paperwork because something felt off. “If I hadn’t [done that],” he told Votebeat, “we’d be still sitting on mistakes.”

18.5 — The Canvass

The canvass — the public meeting at which the results of an election are officially examined and certified — convened at the county elections office on Thursday, March 14. Thirteen precinct judges and a small number of party officials attended. The meeting lasted about thirty minutes.

One by one, the precinct judges called out the corrected totals from their precincts. Some offered brief explanations for what had gone wrong. Poor penmanship that had been hard for them to read back off their own tally sheets. Accidentally writing the wrong numbers. Miscalculations. Joy Smith, the first-time precinct 4 judge, had reported 451 as a proposition total when the correct number was 415 — a digit transposition. “I literally just switched the two numbers when I wrote it down on the form,” she said, attributing the error to exhaustion in the early-morning hours after a long night.

David Treibs, the Fredericksburg Tea Party member who had been the chief public advocate for hand counting, and who had served as the Precinct 13 judge at an auction house, had to correct a total where he had simply forgotten to add two additional ballots to a running count. “There were two ballots, and I just didn’t add them up,” he said in a video posted afterward on Mike Lindell’s social-media platform. “So I would have had to add 450 and two, and it would have been 452 and I didn’t. I just forgot to fill it in.”

After the corrections were called out and the results certified, a small cheer went up in the room. “We did it!” someone said. The precinct judges began to leave. Campbell and the elections staff stayed behind to do the final step: manually entering the official totals into the Texas Secretary of State’s reporting system.

That’s when the second round of errors surfaced. When ballots are scanned using voting equipment, the equipment produces a report in the exact format the state requires — totals separated by Election Day, early voting, and mail-in columns. Campbell’s spreadsheet, which had been organized by precinct and by candidate, did not match this breakdown. To produce the state-required format, the staff would have to re-aggregate the numbers again, pulling out which votes had been cast on Election Day versus early versus by mail.

They printed the spreadsheets. They laid them out across four tables. They read the totals aloud while adding them on a calculator, wrote the new aggregated numbers onto fresh sheets, and then manually entered those numbers into the state system.

The process took nearly three hours. It took that long not because the arithmetic was complex — it was elementary addition — but because the staff kept catching themselves making new data-entry errors as they read the corrected totals out loud, wrote them down, and entered them into the state system. Every handoff produced the possibility of a new transposition or a new missed sum. An hour after Campbell had formally certified the results, he noticed that the numbers for one proposition still did not add up. He had to call the early voting ballot board chair, who had already gone home thirty minutes away, and ask her to drive back to the elections office to fix one more discrepancy.

“It’s my mistake for not catching that,” he said to the reporter still in the room. “I can’t believe I did that.”

Campbell continued to consider the effort a success. The errors, in his view, had been found and corrected before the final results left the county. Nothing had changed the outcome of a race. “The votes were counted correctly,” he told Texas Scorecard. “Nowhere was there an error on a tally sheet. The tally sheets were all correct. We had 8,266 voters and accounted for 8,266 voters in the end.” His lesson learned, he told the reporter, was “you’ve got to do the math and double-check the results. You have to add your columns two or three times in every race to account for every voter.”

18.6 — The Layer Where the Errors Lived

Campbell’s account of what went wrong and what did not is structurally unusual in the election-integrity literature. Normally, the person responsible for an election failure is reluctant to localize the failure with any precision — either out of genuine uncertainty about what happened, or out of the political preference for ambiguity. Campbell did the opposite. He named the layer where the errors occurred. He did it in his own voice, under his own name, to a reporter sympathetic to the hand-count cause. And he described the nature of the errors with technical specificity: addition errors, in which the sum of tally-sheet rows was computed incorrectly when transferred to the reconciliation form; and transposition errors, in which correctly-computed sums were written onto the reconciliation form in the wrong order of digits.

The tally sheets, Campbell said, were correct. The reconciliation forms, which were produced by copying and summing from the tally sheets, were where the errors lived.

This framing — which Campbell had every incentive to avoid if it were not true, since it makes his own party’s method look foolish — maps directly onto the three-layer taxonomy this collection uses to describe vote-reporting vulnerabilities. The counting layer is where votes are read off ballots and totaled. The reporting layer is where those totals are transmitted, summarized, transcribed, and eventually entered into the state system. The certification layer is where the transmitted results are formally attested and published as official.

In Gillespie, the counting layer was manual: five-person teams reading ballots aloud and keeping triplicate tallies. The reporting layer was also manual: precinct judges copying tally-sheet numbers onto reconciliation forms, Campbell himself summing them into a spreadsheet, elections staff re-aggregating the spreadsheet into the state format and entering the results by hand. The errors happened, by Campbell’s own account, every time a correctly-produced number crossed one of the boundaries between those layers.

It is worth naming the honest limit here. AV as a methodology addresses what happens in and after the reporting layer; it does not independently verify what happens in the counting layer. Campbell’s claim that the tally sheets were correct is a claim outside AV’s observability. The five-person counting teams could, in principle, have made errors earlier in the process that Campbell’s spreadsheet reconciliation could not have detected — if a team miscounted a batch of fifty ballots and produced an internally-consistent-but-wrong tally, every subsequent step would faithfully propagate the wrong number. Jennifer Morrell, CEO of the Elections Group and an expert in post-election auditing, told Votebeat: “The only way I can think to validate what they claim are the final vote counts is to run all those ballots through a voting system and compare the two outcomes.” Morrell’s point is that the counting layer in Gillespie has never been independently verified. Neither Campbell nor Netherland nor anyone else can state with external certainty that the tally sheets were accurate. That remains an assumption.

This is a feature of the methodology to state clearly rather than hide. AV addresses the transcription chain above the tally sheet, not the count below it. It photographs the last-mile paper artifact produced by the counting layer, and it locks that artifact in as an independent reference against which every downstream transcription step can be checked. Whatever the counting layer produced, AV creates a contemporaneous record of what it said it produced, before that number began the journey through forms, spreadsheets, re-aggregations, and state systems in which — as Gillespie demonstrates — additional errors are plausibly introduced at every step.

18.7 — What AV Would Have Contributed

Applied to Gillespie, AV would have been mechanically indifferent to the choice of counting method. Its observation moment is the same regardless of what is being observed: the last paper artifact produced at the precinct, timestamped and preserved independently before any transcription begins.

In Gillespie’s case, that artifact was the precinct tally sheets — the pages on which the five-person teams recorded the totals they had agreed upon, and which Campbell was confident were correct. A photograph of each precinct’s tally sheets, taken at the end of counting and before the judge began copying totals onto the reconciliation form, would have created a permanent, verifiable record of what the counting layer produced. Every subsequent error Campbell had to chase — the twelve precinct reporting forms with addition and transposition mistakes, the post-certification discrepancy that summoned the early voting board chair back from home, the three hours of re-aggregation errors — was an error in the chain from the tally sheets to the state system. AV’s photograph would have been a fixed point against which that chain could have been continuously checked.

Scott Netherland’s initiative did, in effect, what AV does structurally. He went back to his tally sheets, compared them to what had been reported, and identified the discrepancy. It took him a morning and a drive back to the elections office. It required him to have been both present at the count and willing to question his own work. It depended on his being the kind of person who wakes up uneasy and double-checks. AV removes the dependence on any of those things. The photograph, taken once, serves the same verification function every time a downstream number is questioned — by a candidate, by a reporter, by a party official, by a skeptical voter. It does not require any specific individual to have done the right thing at the right moment.

AV also would not have required the Gillespie Republican Party to trust any external institution in order to verify its own results. The photographs would have been taken by AV personnel at the precincts, with the tally sheets visible and the time stamp unambiguous. The comparison to the reported numbers would have been arithmetic, not judgment: if the tally sheet says 197 for Precinct 6 and the state system says 207, there is a discrepancy. If it says 451 and the state system shows 415, there is not. The evidence does not require trusting an institution; it requires reading a receipt.

The Gillespie case also suggests something about where the reporting layer’s fragility scales. Travis County Republicans, who hand-counted only their mail-in ballots — just under two thousand of them, a small fraction of their total — also had reporting-layer errors in the 2024 primary. They had misreported undervotes as overvotes in a congressional race. They caught the errors, in their case, with help from the Travis County elections department and the Texas Secretary of State’s office, and through a process of sustained quality-checking that was video-recorded end-to-end and took eight additional hours after the initial count. A rigorous hand count, more rigorously audited than Gillespie’s, still produced reporting-layer errors. The fragility is in the layer, not in any particular execution of it.

Two years later, Gillespie County Republicans hand-counted their primary again. Having failed to recruit enough workers for a second full hand count, they restricted hand counting to Election Day ballots — fewer than three thousand ballots instead of eight thousand — and tabulated early and mail-in ballots by machine. Tom Marshall, the Precinct 1 chair, told Votebeat that redesigned tally sheets were supposed to prevent the transcription mistakes of 2024. Still, counting and reconciliation stretched until nearly 3 a.m., and the county did not submit its report to the state until after 5 a.m. Netherland, again the Precinct 6 judge, told Votebeat around 11 p.m.: “We could have been done three hours ago” if the ballots had been scanned. The 2026 repeat forecloses the reading that 2024’s problems were a first-time-inexperience artifact. Two years of practice, redesigned forms, a smaller workload, and the reporting-layer fragility is still there. AV’s value does not diminish with practice; the transcription chain’s error rate does not fall to zero no matter how many times a jurisdiction runs it.

18.8 — What We Know and What We Don’t

What the record establishes: reporting forms from twelve of Gillespie’s thirteen precincts contained errors on the morning after the 2024 Republican primary. The errors were identified by the county GOP chairman himself, using a reconciliation spreadsheet he built over the weekend, after a precinct judge — on his own initiative and without any required double-check — identified a seven-race discrepancy at his precinct and alerted the chairman. The errors were corrected at the public canvass on March 14, after which additional errors were identified and corrected in the re-aggregation step for the state reporting system. The chairman stated publicly that the underlying tally sheets were correct and that errors occurred only in the transcription onto reporting forms.

What the record does not establish: whether the tally sheets themselves were accurate. Campbell asserts they were; no independent audit has been conducted. The speed of the count — approximately fifty ballots per ninety-minute batch, or roughly three seconds per contest examined by a team of five — was significantly faster than peer-reviewed studies of hand-counting suggest is compatible with low error rates. Charles Stewart of MIT, commenting on the Gillespie pace: “Given the speed of this count, I worry about its accuracy.” Cathy Darling-Allen, the elected elections clerk in Shasta County, California, whose own county’s time-and-motion study found that trained staff averaged seventy-five minutes per batch of twenty-five ballots followed by nine minutes of audit: “Without any audit at all, I would say the results have no validity.” Peer-reviewed experiments on hand-count accuracy have found error rates of up to 1% for the standard read-and-mark counting method and up to 2% for sort-and-stack methods, with higher rates for more complex tasks and for counts conducted under time pressure. Whether Gillespie’s specific tally sheets fall inside or outside those ranges cannot be known from available evidence.

What the record also does not establish: whether all the reporting-layer errors were found. Campbell found twelve of thirteen precincts’ worth of errors in his first pass. He found more during the canvass when judges called out corrections. He found another one an hour after certifying. He found still more during the three-hour re-aggregation. At each stage, additional errors surfaced. The pattern — errors identified in the course of trying to correct other errors — does not permit a confident claim that the final numbers submitted to the state were free of error. It permits only the weaker claim that no further errors were noticed before the forms were submitted.

What the record does not establish but is sometimes assumed: that the errors were too small to matter. None of the corrected errors changed the outcome of a contested race in the 2024 primary. But the largest error required to change a race outcome is not the same as the largest undetected error that might still be present. A race that finishes thirty points apart may also contain a hundred-vote reporting error that no one notices because the outcome is not close; a race that finishes one point apart depends for its correctness on every number upstream. In jurisdictions where outcomes are close, the same fragility that produced harmless errors in Gillespie can produce consequential ones.

18.9 — Why It Matters

The case for AV does not depend on any argument about whether hand counting or machine tabulation is preferable at the counting layer. Reasonable people disagree about the tradeoffs, and that disagreement is largely orthogonal to the question AV is trying to answer. The counting layer produces some piece of paper that says how the votes came out at a precinct. Whether that paper was produced by a DS200 scanner or by a five-person team reading ballots aloud, the piece of paper then has to travel through a chain of human transcription before the number it carries becomes an official result. Gillespie demonstrates that this chain — the reporting layer — is fragile independent of what produced the paper. Twelve of thirteen precincts got it wrong in 2024. A redesigned version of the process still had problems in 2026. A better-resourced version of the process in Travis County still had problems the same year. The layer, not the method, is the point of failure.

This is also why AV can hold its position without taking sides in the larger political dispute over voting machines. A voter who believes machines are untrustworthy and wants to count by hand, and a voter who believes hand counting is inaccurate and wants to trust the scanners, share a common interest: once the counting layer has produced a precinct total, both of them would prefer that the total travel unchanged through the reporting chain to the final result. AV provides a receipt that both of them can read. Its photograph of the tally sheet (or the poll tape) does not require agreement on what happened before the tally sheet; it only locks in what the tally sheet said. Every downstream number can be checked against it. Whatever disputes remain about the counting layer — and those disputes will continue — the reporting layer at least is no longer an invisible black box in which correctly-produced numbers can quietly become incorrectly-reported ones.

Gillespie’s 2024 primary was saved from its reporting errors by one election judge’s initiative, one chairman’s willingness to spend a weekend with a spreadsheet, and the thin margins by which the reported errors happened not to change any outcome. None of these safeguards are reliable in aggregate. AV’s function is to turn what saved Gillespie this time — the discovery of a twelve-of-thirteen discrepancy before certification — into a structural property of the reporting layer itself, available in every precinct, every election, whether anyone wakes up uneasy or not.

18.10 — Further Reading

Primary news coverage — Votebeat / Texas Tribune

Other contemporary coverage

Academic and methodological background

Policy analyses

18.11 — Related Cases in This Collection

  • Case 12 — Prince William County, Virginia, 2020. The canonical AV-would-have-caught-this case — a reporting-layer error caused by scanner poll tapes being misaggregated, discovered over a year after certification. Prince William with machine counting, Gillespie with hand counting: the same reporting-layer error class, produced in different ways. The independent precinct-level photograph is the remedy in both cases.
  • Case 13 — Antrim County, Michigan, 2020. A reporting-layer error that happened to favor the opposite party from Prince William’s. Read alongside Gillespie and Prince William, the three cases make the collection’s non-partisanship argument empirically: reporting-layer errors are indifferent to partisan direction, to counting method, and to the identity of the officials running the count. AV’s value is structural, not ideological.