🗳️ Winning is fine, losing is fraud

Flávio Bolsonaro had no complaints about Brazil's electronic voting machines on Oct. 4, when they put him first with 47.03% of valid votes. He discovered their flaws on Oct. 25, after the same machines handed President Luiz Inácio Lula da Silva a fourth term by 51.79% to 48.21%. He says the results are "mathematically impossible" to accept.

Within a day, he had crossed into Argentina, then flown to Miami, where he released a letter calling himself Brazil's "true president-elect" and the runoff a fraud. He has presented no evidence of electoral fraud.

The evidence to check him, however, is public. Brazil's Superior Electoral Court (TSE) publishes results for every polling station in the country, roughly 500,000 of them, each corresponding to a single voting machine, along with registered voters and turnout, within hours of the count. 

Few large democracies make their arithmetic this easy to audit. We ran both rounds through a standard electoral forensics tool: the second-digit Benford test. The 2026 returns pass it, for both candidates, in both rounds.

Counts that occur naturally and span several orders of magnitude, from town populations to river lengths to vote tallies, do not distribute their digits evenly. In the second position, a zero shows up about 12% of the time and a nine about 8.5%, and the average second digit settles at 4.187. Numbers that have been invented, inflated, or nudged tend to drift away from that pattern, because whoever alters them rarely thinks about what the second digit ought to look like. The test does not prove fraud. It shows where the numbers stop behaving like numbers.

Aggregated to the 5,716 municipal units the TSE reports (5,571 municipalities plus 145 overseas locations), the 2026 results behave impeccably. The average second digit ranges from 4.130 to 4.193 across the four series, one per candidate per round. 

None of the four deviates from Benford's expectation by a statistically significant margin, with p-values running from 0.17 to 0.97, far above the conventional 0.05 threshold. Lula's first round fits almost perfectly. Flávio's runoff, the round he calls rigged, sits at 0.53. The weakest fit of the four belongs to his first round, the one he won and celebrated, and even that one is unremarkable.

The picture changes at the polling-station level, where all four series fail the test decisively. That sounds alarming until one looks at the numbers involved. The median candidate count per machine runs between 102 and 124 votes, and most stations cluster in a narrow band of a few hundred voters. 

Benford's law needs data that span across magnitudes, and a single voting machine does not supply them. The method's own architects warned against applying it to individual machines for exactly this reason.

More telling is how the failure is distributed. It hits Lula and Flávio alike, in the first round as much as in the second. A fraud signature that appears for the winner and the loser, in the round the accuser won and the round he lost, is no signature at all. It describes the instrument, not the election.

Brazil offers a rare natural experiment here. Fraud allegations usually come from losers, which makes them hard to separate from the incentive to contest defeat. The Bolsonaro family has a habit of breaking that pattern. Jair Bolsonaro, now convicted of plotting a coup and under house arrest, won the presidency in 2018 and still insisted that fraud had denied him a first-round victory. He went on to attack the voting machines for the rest of his term.

His son's case is cleaner still. The two rounds of 2026 used the same machines, the same voter registry, the same polling stations, and the same counting procedures, two weeks apart. The outcome was the only variable that changed. If the system had been compromised, the traces should appear in both rounds. They appear in neither.

Flávio's specific grievance concerns Justice Cristiano Zanin's decision to extend voting by two hours on runoff day, which he claims added 10 million votes to Lula. Both campaigns had requested that extension — and it ultimately produced fewer than 36,000 extra votes.

More hours produce more voters, and more voters produce larger counts, but none of this alters the digits of a tally. Whatever one thinks of the decision, it is a legal argument, not a counting anomaly, and one his campaign endorsed before it lost.

Benford's law is a smoke detector, not a fire investigator. Political scientists have shown it can flag clean elections and stay silent on dirty ones, and its defenders answer that it was only ever meant to tell analysts where to look. We side with the defenders, and in 2026 it tells us there is nothing to look at. Manipulation subtle enough to leave every digit in place would slip past it.

What the test does rule out is tampering on the scale Flávio's story requires. Overturning a 4.4-million-vote margin while leaving the second digits of 5,700 municipal counts precisely where Benford's law expects them would demand a precision no known fraud scheme has ever achieved. Anyone who claims it happened owes Brazilians an explanation of how it was done without leaving fingerprints.

Winning is fine, losing is fraud. The numbers do not share that philosophy.

Dalson Figueiredo is an associate professor of political science and a CNPq research fellow. 

Ernani Carvalho is a full professor of political science, dean of the Graduate Program in Public Policy at the Federal University of Pernambuco (UFPE), and a CNPq research fellow. 

This article is the second in a series by The Transparent Ballot Box Project, an initiative of Transparência Eleitoral that applies forensic methods to the 2026 presidential returns. Replication materials are available here.