INDEPENDENT RESEARCH · PUBLIC DATA

More context.
Better questions.
Open data.

A look at how Brazilian elections work. Explore operational indicators, compare results and see the limits of each analysis.

Explore the study ↗How it works →

Your contribution helps improve the hardware and expand data processing. Voluntary support, no minimum amount.

EDITION OF 2026-10-242022 — 2026

Charts and PDF documents are in Portuguese; this page explains them in English.

SUPPORT THE RESEARCH · PIX

Help expand
the next study.

Voluntary contribution, no minimum amount, used to improve the hardware and data-processing capacity. Goal for improvements: R$ 5.000. Pix is Brazil's instant payment system.

PIX KEY (RANDOM)4cc4708e-e291-4012-ab41-3c67eab901c9

Pix shows the account holder: Leandro de Almeida Miguel. Check the recipient in your banking app before confirming. Contributing is optional; access to data and documents is free.

Pix QR code provided by the project lead. Recipient: Leandro de Almeida Miguel. — Point your banking app's camera at the QR code

Transparency from the start. Results are from the edition of 2026-10-24, revised after an external review (see the change log in the technical document, in Portuguese). Indicators do not prove irregularities and do not reveal how late or manually enabled voters voted.

See limitations ↗

11 statesCovered by the study

3 years2022, 2024 and 2026 · five rounds

385 thousandPolling sections in the 2026 first round

29,032Polling places with flagged behaviour analysed

01 / EXPLORE

Numbers need
context.

Pick an election round and examine the indicators of the current edition. Operational categories may overlap.

OPERATIONAL INDICATORS

—

EDITION 2026-10-24

Eligible sections

With some flag

Share flagged

Late voting or late closing

Manual voter enablement above the defined thresholds

0Scale relative to the largest indicator of the round

MAIN RESULT · SAME BUILDING (3 OR MORE NEIGHBOURING SECTIONS WITHOUT FLAGS)

Flagged sections comparable within the building

Above the thresholdunflagged:

Of those, in favour of side Aunflagged:

Side A = PT (president) or the leading mayoral candidate in the municipality. "Above the threshold" = an absolute difference of 5 percentage points or more that also exceeds the normal variation between rooms of the same building (|z| ≥ 2). Another flagged sections have no comparable neighbour and are compared with their municipality, separately (see the technical document).

A flag means the section reached one of the operational thresholds defined in the study. It does not reveal the choice of anyone who voted late or was enabled manually.

01b / LOOK UP POLLING PLACES

Find a polling
building.

Choose a state and type part of the municipality name, the polling-place name or the electoral zone number. The list shows polling places with flagged behaviour in at least one election, with observation priority, points and sections to check. Codes and names are the official ones from the TSE (Brazil's electoral court); places without a registered name are shown as such.

Choose a state to start.

Municipality	Zone	Place	Place name	Priority	Points	Result signal	Flagged sections	Rec. 2+	Pres. dif.	Direction	Sections to check	Manual enablements 2026 R1 (unassoc.)	Field kit

Manual enablements 2026 R1: voters enabled without biometrics in the 2026 first round, summed over all sections of the building; in brackets, how many and what share were authorised by a fingerprint that the voting machine does not associate with a registered poll worker. Reference: 39% in a random sample of 5,000 sections. Values above that usually mean a poll worker whose fingerprint was not registered at opening: a traceability gap, not an improper vote. Computed for high-priority polling places; the field kit (PDF, in Portuguese) lists the sections to check and the observation steps for election day.

02 / CHARTS

Twelve charts,
each with "how to read".

The same charts as in the chart notebook (labels in Portuguese). Click to enlarge. The numbers come from the files in the Open data section.

Chart 1: Sections with flagged behaviour, by election — **Chart 1 · Sections with flagged behaviour, by election**First rounds have between 367 and 385 thousand sections analysed in the 11 states; mayoral second rounds only where they took place (77 thousand). Late voting is much more common in presidential first rounds, consistent with longer queues, than in mayoral elections; excessive manual voter enablement appears in similar numbers in all of them, because it is defined by the 99th percentile of each state. See in the notebook (PDF, Portuguese) ↗

Chart 2: Difference above the threshold, by comparison level — **Chart 2 · Difference above the threshold, by comparison level**Difference above the threshold = at least 5 percentage points, in absolute value, in the share of side A, and |z| ≥ 2. The orange bar is the same test on sections without flagged behaviour: what is expected by chance and normal variation. Under each election, the number of sections in each bar. On the left, the main result: in the same building, in 2026, 664 of 15,593 flagged sections (4.3%) exceed the threshold, against 7,175 of 293,231 unflagged ones (2.4%): a ratio of 1.7. On the right, sections without a comparable neighbour (a single section or a whole building flagged): the reference is the municipality, the rate is several times higher in both groups (7.8% against 13.2% in 2026), and that is why this level is kept separate. Bars for the mayoral second round have few sections. At any level, most flagged sections do not exceed the threshold against their reference, which does not mean identical voting: it means a difference below 5 points or within normal variation. See in the notebook (PDF, Portuguese) ↗

Chart 3: Direction of the differences, by level — **Chart 3 · Direction of the differences, by level**Side A = PT (president) or the leading mayoral candidate in the municipality; the 50% line marks an equal split between A and B, not an expected neutrality: the relevant reference is the orange bar of unflagged sections. Under each election, the number of differences above the threshold in each bar. In the same building, in 2026, 76% of the differences of flagged sections lean towards PT, against 54% among unflagged ones; in the 2022 first round the gap is small (63% against 57%). In mayoral elections there is no consistent direction, and the bases are small. At municipality level both groups lean towards A, which shows why mixing levels inflates the reading. Direction is not cause: it may reflect the profile of those who face queues. See in the notebook (PDF, Portuguese) ↗

Chart 4: Direction test by permutation — **Chart 4 · Direction test by permutation**Each point is the voting difference (side A minus side B) between flagged and unflagged sections, measured as a deviation from the mean of the polling place itself and weighted by valid votes. The bands run from the 2.5th to the 97.5th percentile of the differences obtained by shuffling the flags 500 times: the wide one within the municipality; the thin dark one within the polling place itself (closer to the same-building question). A point outside the bands = a difference larger than the shuffling produced 95% of the time, under the hypothesis of flags at random; it is not a measure of importance or cause. Both schemes give the same verdict in every case. For mayor in 2024 no point leaves the band. Turnout, cancellations and the combined scores do not define the behaviour used in the scan and appear only in the table of the technical document. See in the notebook (PDF, Portuguese) ↗

Chart 5: Polling places by state and priority — **Chart 5 · Polling places by state and priority**Above each bar, the total number of polling places and, in brackets, how many are high priority. BA and PE account for 287 of the 451 high-priority polling places. See in the notebook (PDF, Portuguese) ↗

Chart 6: Distribution of points — **Chart 6 · Distribution of points**The typical case is a single flagged section in a single election (43% of polling places). The 25- and 45-point cut-offs separate medium and high priority. The vertical scale is logarithmic so that the tail shows. See in the notebook (PDF, Portuguese) ↗

Chart 7: Composition of points (high priority) — **Chart 7 · Composition of points (high priority)**Averages over the 451 high-priority polling places; the three components that weigh most are recurrence (2+ elections), sections flagged in the 2026 first round, and the volume of late/manual votes. On average, 32 points are operational (recurring queues/manual enablement, volume, behaviour in 2026) and 22 come from results. 120 of the 451 have no section with a difference above the threshold and 28 have no result points at all: they enter through operations, not results. See in the notebook (PDF, Portuguese) ↗

Chart 8: Mayor × president, same sections — **Chart 8 · Mayor × president, same sections**The same sections in the three first-round elections, with explicit denominators. In the right panel, the "same building" restriction is applied election by election, so denominators change between offices. Among flagged sections, the share above the threshold is of the same order for both offices (slightly higher for mayor in the all-sections panel, not in the same-building one). The raw count of sections deviating "only in presidential elections" is much larger because there are many more sections with queues in presidential first rounds, not because they deviate more. Strict comparison, sections flagged for both offices (President 2026 R1 × Mayor 2024 R1, 1,445 sections): 6.1% against 6.4% (McNemar p 0.74): no evidence of asymmetry between offices. See in the notebook (PDF, Portuguese) ↗

Chart 9: Recurrence across elections — **Chart 9 · Recurrence across elections**The vast majority appears only once. Among sections flagged in the 2022 first round, 37% were flagged again in the 2026 first round, against 6% of those that were not (6.0 times more). Recurrence may have structural causes (a large building, voters without biometrics, a zone that always has queues); that is why it counts in the score, without indicating any irregularity. See in the notebook (PDF, Portuguese) ↗

Chart 10: Votes after 5 p.m. — **Chart 10 · Votes after 5 p.m.**Direct count in the voting-machine logs: "vote computed" events timed after 5 p.m. on election day. In presidential first rounds they number in the millions; in second rounds and municipal elections, far fewer. Voting after 5 p.m. is legal for those already in the queue; the number measures queues, not irregularity. See in the notebook (PDF, Portuguese) ↗

Chart 11: Observation map by state — **Chart 11 · Observation map by state**The map covers 2,806 polling places and 14,107 sections in the 11 states; the suggestion is 1,826 observers (1 per high-priority place and 1 for every 2 medium-priority places, rounded up within each zone). 861 of these places have at least one section with a result difference above the threshold. See in the notebook (PDF, Portuguese) ↗

Chart 12: Cities with the most high-priority polling places — **Chart 12 · Cities with the most high-priority polling places**Large cities appear because they have more polling places; to allocate observers, what matters is the list by zone and place within each city, in the observation map. See in the notebook (PDF, Portuguese) ↗

03 / HOW TO READ

An observation.
Not an accusation.

The study combines operational records with aggregated results per polling section. The aim is to raise verifiable questions and support further observation. The score of each polling place is an order of priority for observation, not a probability of irregularity.

Read the technical document (Portuguese) ↗

Public sources

Urn tally reports (BU), voting-machine logs, machine replacement records and voter-profile data, from the TSE Open Data Portal.

Explicit references

The main result compares each section with its neighbours in the same building. Sections without a comparable neighbour are compared with their municipality and kept separate, because the base rate there is different.

Limits in plain sight

Queues, biometrics, infrastructure and the make-up of the electorate can explain the differences. Association does not establish cause. How late voters voted is unknown.

04 / LIBRARY

Read. Check. Question.

Documents to follow the reasoning and assess the results, in Portuguese. Edition of 2026-10-24.

↗START HERE · PDF (PT)

The study
in plain words

An introduction to the question, the scope and the limits.

Plain summary 03 pages ↗VISUALISATIONS · PDF (PT)

Charts to
explore the patterns

Twelve charts with reading guidance.

Chart notebook 12 pages ↗IN DEPTH · PDF (PT)

Sources, criteria
and procedures

Definitions, coverage, tests, scoring, limits and change log.

Method and data 07 pages

↗REPORT · PDF (PT)

Scan by
polling place

Tables by election, comparison level, polling place and zone.

Scan 12 pages ↗OPERATIONAL · PDF (PT)

Observation map
2026 second round

Overview of the states; maps by state in the data package.

General map 07 pages ↗STUDY · PDF (PT)

Indicators per section
and permutation tests

Eight indicators, direction tests and an observation tool.

Inconsistencies 12 pages

↗NOTE · HTML (PT)

extensão da votação até as 19h no 2º turno de 2026

Note generated from the data, in Portuguese.

Data note NOTA_HORARIO_19H_2026T2

05 / OPEN DATA

The analysis doesn't
have to end here.

Organised datasets to query, reproduce and propose improvements without repeating the whole collection. CSV files use a semicolon separator and a comma as the decimal mark; column names are in Portuguese and a field dictionary is included. Everything derives from public TSE data and is licensed CC BY 4.0: use, adapt and redistribute citing the source and keeping the notice that a flag is not an accusation.

Download summary by election ↧

Code and reproduction: https://github.com/Leandro-VPN/observatorio-do-voto. The raw data (urn tally reports, voting-machine logs, machine records and voter profile) are public at the TSE and are downloaded by the repository scripts.

CSV

Summary by election and round

Aggregated indicators by comparison level

2 KB↧

CSV

Comparison by level

Flagged × unflagged, by election and level (place / few pairs / municipality)

2 KB↧

CSV

Mayor × president, same sections

Rates with denominators; see also pres-x-pref-mesmas-secoes.csv (McNemar)

1 KB↧

CSV

Mayor × president, strict comparison

Same sections flagged for both offices; McNemar test

0 KB↧

CSV

Conditional recurrence

Probability of repeating the behaviour across elections, among sections present in both

0 KB↧

CSV

Direction tests by permutation

Two shuffling schemes, null quantiles, p and q

6 KB↧

CSV

Coverage and indicators by election

Sections, logs, votes after 5 p.m., manual sequences, counts per indicator

1 KB↧

CSV

Field dictionary

By file and column: type, % missing, example and description (in Portuguese)

78 KB↧

CSV

Methodological glossary

Definitions of behaviour, comparison, threshold, scoring and priority (in Portuguese)

8 KB↧

How to use the data

File format and sample DuckDB queries, including joins that do not double-count turnout (in Portuguese)

2 KB↧

CSV

Field kit: sections of high-priority polling places

One row per section: last vote and closing history, manual enablements by type (2026 R1), what to observe and blank columns for notes on the day

1,0 MB↧

ZIP

Scan by polling place

Places with score and reasons, 2026 R2 priority, history by section, sections by election · GitHub Release dados-2026-10-24-r2

16,1 MB↧

ZIP

Observation map 2026 R2

Places, checklist per section, zones; general and per-state PDF and Excel · GitHub Release dados-2026-10-24-r2

6,1 MB↧

ZIP

Scan spreadsheets

VARREDURA_LOCAIS.xlsx and LOCAIS_<UF>.xlsx · GitHub Release dados-2026-10-24-r2

83,1 MB↧

PARQUET