Public Sector · Election Administration
VANTAGE finds voter records that belong to the same person, even when the name, address, or spelling doesn't match exactly. Election staff get a scored list of likely duplicates to review, so the list stays accurate and no eligible voter is removed by mistake.
Voter lists change every day. People move, marry, update a driver's license, or register again through a different office. Each of those events can leave a second record behind.
Researchers and election officials have documented the same pattern for years: registration data goes stale and splits into duplicates as people and agencies update it.
About 24 million U.S. voter registrations were no longer valid or had significant inaccuracies, according to a 2012 Pew study.
People with registrations in more than one state, found in the same Pew study of registration records.
Records one state's motor vehicle agency inadvertently duplicated ahead of a 2018 primary, as reported by the MIT Election Data and Science Lab.
Cases in a single county where one voter was given a new ID, splitting one person into two records (MIT Election Data and Science Lab, 2019).
The Help America Vote Act asks states to keep a statewide list in which "duplicate names are eliminated," and also to keep "safeguards to ensure that eligible voters are not removed in error." VANTAGE is built around both requirements: find the duplicates, and give staff the evidence to confirm each one before anything changes.
Sources: Pew, Inaccurate, Costly, and Inefficient (2012) · MIT Election Data and Science Lab (2019) · 52 U.S.C. § 21083
Every list goes through the same documented process, and every result arrives with the evidence behind it.
The file is treated as read-only the moment it arrives. VANTAGE checks file integrity, puts every record into a common voter record layout, and scores each record and field for completeness and consistency. Problems are graded Critical, Warning, or Informational, and a quick pre-screen gives an early read on how large the duplicate problem is.
High-certainty identifiers are compared under a hierarchy of match rules. The answer is yes or no, so these pairs go straight to the top of the queue. Pairs that span two counties are flagged first, because no single county office can see them on its own.
Addresses are broken into their parts before they are compared. Names are compared by sound, spelling, and meaning, which catches nicknames, transliterations, compound surnames, and transposed birth dates. A statistical model then estimates how likely each pair is the same person. It calibrates on your own data, so no hand-labeled training set is needed.
Pairs in the middle of the range go to an AI reasoning layer that weighs the whole record and writes a plain-English explanation for any change to the score. Every adjustment is documented, so reviewers and auditors can see why a pair moved.
Running alongside matching, this scan flags identifiers that contradict each other, birth dates that don't fit the registration date, addresses outside the reported county, and placeholder or test entries. Anomalies are reported on their own and are never a finding about any voter's eligibility.
Results from both passes are combined into one scored list and sorted into four bands, so staff spend their time where it matters most.
| Band | Score | What it means | Recommended handling |
|---|---|---|---|
| Confirmed | 95 to 100 | Same person, supported by strong agreement across identifiers | Ready for resolution under the office's existing list maintenance procedures |
| High Confidence | 80 to 94 | Very likely the same person, with minor differences explained | Staff review using the field-by-field evidence provided |
| Medium Confidence | 50 to 79 | Possible match that needs context to decide | AI-assisted review with written rationale, then staff decision |
| Low Confidence | Below 50 | Shared traits, but likely two different people | Kept for reference and generally treated as distinct voters |
Each pair also lists the specific fields that drove the score, so a reviewer never has to guess why two records were linked.
These are the screens election staff use every day, shown here with made-up sample data. Pick a screen to see how a list moves from upload to resolved.
Analytics
Sample jurisdiction · 120,000-record voter list
Duplicates and errors
AI-flagged duplicate voter records from the selected upload
sample_voter_list_2026.csv ▾| Pair ID | Name | County | Score | Band |
|---|---|---|---|---|
| HB10421 | MARIA LOPEZ GARCIA | Harbor | 98.67 | CONFIRMED DUPLICATE |
| HB10988 | MARIA LOPEZ-GARCIA | Harbor | 98.67 | CONFIRMED DUPLICATE |
| CR22017 | DANIEL MOORE | Cedar Ridge | 93.10 | HIGH CONFIDENCE |
| CR22740 | DANIAL MOORE | Cedar Ridge | 93.10 | HIGH CONFIDENCE |
| MB05112 | KAREN HILL | Millbrook | 97.40 | CONFIRMED DUPLICATE |
| PV31876 | KAREN ASHBY-HILL | Pine Valley | 97.40 | CONFIRMED DUPLICATE |
| LS40311 | WILLIAM GRANT | Lakeshore | 88.25 | HIGH CONFIDENCE |
| LS40957 | BILL GRANT | Lakeshore | 88.25 | HIGH CONFIDENCE |
Pair HB10421 : HB10988 CONFIRMED DUPLICATE
Field-by-field comparison of every record in the duplicate group
| Field | Record 1 | Record 2 | Match |
|---|---|---|---|
| First name | MARIA | MARIA | Exact |
| Last name | LOPEZ GARCIA | LOPEZ-GARCIA | Normalized |
| Nickname | (blank) | (blank) | Missing |
| Date of birth | 03/14/1986 | 03/14/1986 | Exact |
| Street address | 418 MAPLE AVE | 418 MAPLE AVENUE | Normalized |
| City / ZIP | HARBOR CITY 00412 | HARBOR CITY 00412 | Exact |
| Driver's license | •••• 7731 | •••• 7731 | Exact |
| ID number (last 4) | •••• | •••• | Exact |
| Voter ID | 000001048213 | 000001093377 | Differs |
Anomalies
Voter records flagged with data-quality issues, grouped by status
| ID | Name | DOB | County | Invalid field |
|---|---|---|---|---|
| HB20871 | JAMES PORTER | 02/11/1949 | Harbor | reg_date |
| CR18004 | ELENA BROOKS | 07/22/1991 | Cedar Ridge | id_last4 |
| MB09455 | THOMAS REED | 11/30/1958 | Millbrook | zip |
| PV12930 | AISHA COLEMAN | 05/09/2004 | Pine Valley | dob |
| SB03318 | ROBERT VANCE | 09/01/1940 | Stonebridge | reg_date |
| LS27716 | NINA FARRELL | 12/17/1977 | Lakeshore | id_last4 |
Upload data
Upload a CSV to run field-level validation and duplicate matching
Data loads
Every voter list processed by the matching pipeline
| Run ID | Source | Records | Duration | Status |
|---|---|---|---|---|
| 7f3c…a91e | sample_voter_list_2026.csv | 120,000 | 38.4s | Completed |
| 2b8d…04c7 | sample_voter_list_q2.csv | 118,450 | 37.9s | Completed |
| c41a…77f0 | sample_pilot_extract.csv | 25,000 | 8.1s | Completed |
Administration
Accounts provisioned to use the matching pipeline
| Name | Role | Record allowance | Remaining | Status |
|---|---|---|---|---|
| Avery Collins | Admin | Unlimited | Unlimited | Active |
| Jordan Ellis | Reviewer | 150,000 | 150,000 | Active |
| Morgan Price | Reviewer | 150,000 | 112,400 | Active |
| Riley Nguyen | Reviewer | 150,000 | 150,000 | Active |
| Casey Patel | Analyst | 150,000 | 64,980 | Active |
Sample data only. Names, counties, IDs, and figures are fictitious.
A common duplicate check is a search by name and date of birth. It works when both records were typed the same way. It misses a hyphen dropped in one record, a married name added to another, or a single mistyped letter.
In a 150,000-record pilot, more than half of the duplicate pairs VANTAGE found were caught only by probabilistic scoring. A name and date-of-birth search would have missed them.
Hyphenated and unhyphenated surnames
Married and maiden names
Typos and transcription errors
Compound and culturally specific names
Nicknames and phonetic spellings
Results arrive in formats administrative staff already use, with no special data tools required.
VANTAGE is built for sensitive government records from the ground up.
Multistate data-sharing programs help states find voters who moved across state lines. The duplicates created when people move between counties in the same state need their own solution.
Built for in-state record linkage, where county systems often can't see each other's records.
The model calibrates on the list it receives. Offices never have to label example duplicates.
Handles transliterations, compound surnames, and naming conventions that rules and phonetic codes miss, with a written rationale.
From secure intake to final briefing, Stellar runs the process. No in-house data science team is needed.
A repeatable, documented check across the full statewide list.
Scored pairs to review instead of manual searches.
Dispositions and reports that fit existing procedures.
A clear view of where and why records split.
A pilot runs six to eight weeks from receipt of the data. A typical sample is up to 150,000 records from at least five counties, ideally including both urban and rural areas, delivered as CSV, pipe-delimited text, or a direct database extract.
Sign the NDA and data sharing agreement, agree on a secure transfer method, and set up the processing environment.
Ingest the sample and deliver the data quality report.
Run both matching passes, AI-assisted review, and the anomaly scan, then deliver the scored pair list and anomaly report.
Deliver the final report and brief leadership on findings and recommended next steps.
After a pilot, VANTAGE can run across the full statewide list and repeat on a quarterly, twice-yearly, or annual cycle, with an online dashboard and reporting for state and county staff.