9 — Doing a Comparison Analysis
9.1 — Introduction: What a Comparison Analysis Is
A comparison analysis takes independently-captured poll tape totals and compares them against officially reported election results to determine whether they match. This is the core analytical act that gives Actual Vote its value. Everything else in the system — the app, the archive, the vetting process, the access methods, the volunteer network — exists to make this comparison possible. The comparison itself is where the evidence is produced.
The concept is simple: a precinct tabulator prints a tape showing that Candidate A received 542 votes. The official results report that Candidate A received 542 votes at that precinct. Do the numbers match? If they do, that is positive evidence that vote reporting was accurate at that precinct for that contest. If they don’t, something happened between the poll tape and the official results that needs to be understood.
The concept is simple, but the execution is not. Vote reporting in the United States is not standardized. Each state has its own rules. Many counties have their own conventions within those rules. The path from poll tape to official result involves jurisdiction-specific logic — how precincts map to reporting units, how voting modes are categorized, how split precincts are handled, how adjustments and corrections are applied — and much of that logic is undocumented. A comparison analysis that doesn’t account for these details will produce false discrepancies: apparent mismatches that are actually artifacts of the analyst’s incomplete understanding of the reporting logic.
This section is a practical guide to conducting a comparison analysis. It is written for two audiences. The first is independent analysts — journalists, academics, civic organizations, or individual citizens — who want to conduct their own comparison analysis using the public AV archive and publicly available official results. The second is readers who want to understand in detail how America Counts conducts its analysis.
Both audiences follow the same logical steps. AC’s high-throughput pipeline — AI-assisted transcription, automated Python pipelines, systematic coverage prioritization — automates and scales what an independent analyst does by hand. This section describes the process common to both, notes where AC’s pipeline diverges, and provides practical guidance for anyone who wants to do this work.
9.2 — Scoping Your Analysis
Before you begin the comparison analysis itself, you need to make four decisions that determine its scope and feasibility.
Choose your purpose
Why are you doing this analysis? The answer determines everything else. You might be verifying a specific contest in a specific county because you want to confirm that the reported results are accurate. You might be auditing an entire county’s election as a demonstration project. You might be investigating a tip or a statistical anomaly that someone else noticed. You might be building a statewide dataset as part of a research project. Each purpose implies a different scope, a different level of rigor, and a different amount of work.
For a first analysis, a focused scope is recommended: one county, one or two contests, using the AV archive submissions available for that county. You will learn more about the process from completing a small analysis than from getting halfway through an ambitious one.
Choose your geography
County is the natural unit of analysis. Official election results are typically reported at the county level, and the vote reporting logic that governs how precinct-level results are aggregated into official totals operates at the county or state level. Pick a county — or a small set of counties — to analyze. If you’re working with the AV archive, choose a county where substantial submissions are available.
Choose your contests
You do not have to transcribe every contest on every tape. Poll tapes for a general election may list a dozen or more contests: president, U.S. Senate, U.S. House, governor, state legislature, county commission, school board, judicial retention, ballot measures. Transcribing all of them for all precincts is substantial work.
For a focused analysis, choose the one to three most important races — typically the national or statewide contests where the stakes are highest and the public interest is greatest. For a comprehensive analysis, transcribe all contests. For a targeted investigation, transcribe the specific contest or contests where you have reason to look.
AC’s pipeline prioritizes the most important races first: national, statewide, and competitive contests. This is a reasonable allocation of limited resources toward maximum impact. An independent analyst should follow the same logic — unless their purpose calls for something different.
Estimate the effort
A single county in a single election might have anywhere from a dozen to several hundred precincts. Each precinct has one or more poll tapes. Each tape has multiple contests. Transcribing a full county for all contests is a project that could take days or weeks. Transcribing two or three contests from a dozen precincts is an afternoon’s work.
Be realistic about the time commitment. A first analysis will take longer than subsequent ones, because you will spend significant time understanding the jurisdiction’s vote reporting structure. That knowledge accumulates, and the second analysis of the same jurisdiction will go much faster.
What you’ll need
To conduct a comparison analysis, you will need:
- Access to the AV archive (av.democracycounts.org). This is the source of vetted poll tape video recordings.
- Official election results from the relevant jurisdiction. These are typically available from the county elections office website, the state elections board, or the Secretary of State.
- A spreadsheet application (Google Sheets, Excel, or equivalent) for transcription and comparison. For larger analyses, familiarity with Python (particularly the pandas library) is helpful but not required.
- Understanding of the vote reporting structure for your jurisdiction. How do precincts map to reporting units? Are there split precincts? How are early voting and mail-in ballots reported? This information may be available from the elections office website, state election law, or the canvass report — but often it must be inferred from the data itself.
- Time and patience. This work is genuinely tedious at points. The reward is evidence that currently exists for almost no American election.
9.3 — Step 1: Obtaining the Poll Tape Evidence
From the AV archive
All vetted Actual Vote submissions are available in the public archive at av.democracycounts.org. America Counts reviews every submission and rejects any that are not legitimate poll tape recordings — submissions that are blurry, that show the wrong document, that are incomplete, or that cannot be verified as authentic poll tapes. What remains in the archive has passed a quality gate: these are recordings that AC has determined, through its vetting process (described in Section 8.5), to be genuine poll tape videos.
To begin your analysis, filter the archive to your target: state, county, and election date. Depending on the archive’s interface, you may be able to browse by these fields directly.
You can transcribe directly from the website — watching each video and recording the values in your spreadsheet — or download the videos for offline work. Direct transcription from the website is usually sufficient for small analyses. Downloading is helpful if you expect to re-watch videos multiple times or if your internet connection is unreliable.
Coverage realities
The archive will rarely contain recordings from 100% of precincts in any county. AV depends on volunteers recording poll tapes at their polling places, and volunteer coverage is inherently uneven. Some precincts will have recordings; many will not. For the counties and elections that AC has analyzed, coverage has ranged from a small percentage to substantial but never complete.
Partial coverage is still valuable. Every comparison is a piece of evidence, whether you have recordings from 10 precincts or 100. The strength of the evidence scales with coverage — a county where 80% of precincts match is stronger evidence of reporting accuracy than a county where 10% match — but even 10% coverage produces something that did not previously exist: independent evidence about the accuracy of the reported results at those precincts.
Note which precincts you have and which you’re missing. Incomplete coverage affects what you can conclude about the county as a whole, but it does not invalidate the conclusions about the precincts you do cover.
9.4 — Step 2: Transcribing the Poll Tapes
What transcription means
Transcription is the process of watching each poll tape video and recording the vote totals for each contest you are analyzing. The poll tape is a printout from the precinct’s tabulator — it shows, for each contest, the number of votes counted for each candidate (and typically the total ballots cast, overvotes, undervotes, and other summary statistics). Your job is to transfer the relevant numbers from the video into your spreadsheet accurately.
For each tape, record: the precinct identifier (as printed on the tape), the contest name, each candidate’s name, and each candidate’s vote total — for each contest you are analyzing. If you are doing a focused analysis of one or two contests, you only need to transcribe those contests. If you are doing a comprehensive analysis, transcribe everything.
Best practices for transcription
Use a consistent spreadsheet structure. Before you begin, decide on your column layout and stick with it. A common structure has one row per candidate per contest per precinct: columns for precinct identifier, contest name, candidate name, and vote total. An alternative structure has one row per contest per precinct, with candidates in separate columns. Either works; consistency matters more than format.
Record exactly what the tape shows. This is the most important rule. Do not interpret, round, adjust, or “fix” what you see on the tape. If the tape says 542, write 542. If the tape says “REP - JOHN SMITH” and the official results say “John Smith (R),” record exactly what the tape says — you will reconcile the naming differences in the merge step. Faithful transcription protects the integrity of your data.
Note the precinct identifier as printed on the tape. This is critical for matching your transcription to the official results later. Tapes usually print a precinct number, name, or code somewhere near the top. Record it exactly as printed — including any leading zeros, abbreviations, or formatting. The tape might say “PCT 042” while the official results say “Precinct 42” or “42.0” or “Eagle Creek.” You will need to build a crosswalk between these identifiers in the merge step, and you can only do that if you have recorded the tape’s identifier faithfully.
Flag ambiguities. If a value is hard to read — the ink is faint, the video is blurry at a critical moment, the digit could be a 3 or an 8 — flag it. Use a consistent notation: a separate column for confidence flags, or a color code, or an asterisk. Do not guess. A flagged value can be revisited; a silently guessed value can produce a false discrepancy or a false match.
Watch for multiple tabulators. Some precincts have more than one tabulator (scanner). Each tabulator prints its own tape. If you have only one tape from a two-tabulator precinct, your transcription will undercount the precinct’s total — and this is expected, not a discrepancy. Look for any indication on the tape of tabulator number: some tapes print “Machine 1 of 2” or a tabulator serial number or ID. Note it in your spreadsheet. You will need this information in the aggregation step to determine whether you have all tapes for a given precinct.
Watch for tape types. Some tabulators print multiple tapes over the course of election day: a zero tape at opening (confirming the machine starts with zero votes), a results tape at closing, and sometimes a cumulative or summary tape. You want the closing results tape — the one printed after polls close, showing the final count. Zero tapes will show all zeros; cumulative tapes may show intermediate totals. Make sure you are transcribing the right tape.
Watch for tape sections. A single results tape may have multiple sections: Election Day results, Early Voting results (if the precinct tabulator also processed early votes), and summary totals. Make sure you understand which section you are transcribing and how it relates to the official results. If the official results report Election Day and Early Voting separately, you want the Election Day section of the tape. If the official results report them combined, you may need to sum the tape’s sections.
AC’s transcription approach
America Counts uses AI-assisted transcription — optical character recognition and video analysis — to process poll tape submissions at scale. The AI reads the tape and produces an initial transcription, which is then checked by a human reviewer against the source video. The AI output is a first pass, not the final word. When the AI reads a value with low confidence, when the video quality makes automated reading unreliable, or when the tape layout is unusual, a human transcriber reviews the source video and makes the final determination.
AC may skip or defer blurry recordings, blurry values, or low-quality submissions in initial passes and return to them later. The priority is producing the most high-impact comparisons first — clear recordings from important contests — rather than processing every submission sequentially. This is a triage decision: the clearest recordings of the most important races first, with diminishing-quality material addressed as resources allow.
9.5 — Step 3: Obtaining and Understanding the Official Results
Where to get official results
Official election results are published by the jurisdiction that administered the election. For most elections, this means the county elections office, although state-level aggregation is also available for statewide contests. Common sources include:
- County elections office website. Most counties publish precinct-level results on their website after the canvass is complete. The format, level of detail, and timeliness vary enormously.
- State elections board or Secretary of State website. Most states publish statewide results that include county-level and sometimes precinct-level breakdowns. These are often the most standardized source, but may lack the precinct-level detail needed for a comparison analysis.
- Third-party aggregators. Some organizations (including the MIT Election Data + Science Lab, OpenElections, and state-specific projects) aggregate and standardize official results. These can be useful for cross-jurisdictional analyses but should be verified against the original source.
Always use the certified results if available — the post-canvass, officially certified results that include provisional ballot adjudication, late-arriving absentee ballots, and any corrections made during the canvass. Election-night preliminary results are available sooner but are subject to change. For a preliminary check — a quick look at whether your analysis is in the right ballpark — election-night results can be useful, but the definitive analysis should be against certified results.
Understanding the format
Official results come in many formats: CSV files, Excel spreadsheets, PDF documents, HTML tables on websites, and sometimes only as images or scanned documents. The format determines how much work is needed to get the data into a structure you can compare against your transcription.
Machine-readable formats (CSV, Excel) are vastly preferable. You can load them directly into a spreadsheet or Python script. PDF and HTML require manual extraction or specialized tools. Scanned images are the worst case — they require OCR or manual transcription of the official results themselves, introducing a second source of potential transcription error.
Before you begin the comparison, study the official results to understand their structure. Identify the following:
What is the reporting unit? Are results reported at the precinct level? By precinct group, ward, or district? By some other unit? If the reporting unit is larger than a precinct, you will need to aggregate your tape-level data to match.
Are results reported by voting mode? Some jurisdictions report Election Day, Early Voting, Mail/Absentee, and Provisional results separately. Others combine them into a single precinct total. The poll tape at the precinct typically captures only the votes counted by that tabulator on election day. If the official results combine all voting modes, the official total will be higher than your tape transcription — and this is expected, not a discrepancy. You need to identify whether separate Election Day results are available.
How are precincts identified? By number, name, code, or some combination? Does the identifier in the official results match what’s printed on the poll tape? If not, you will need to build a crosswalk — a mapping between the two sets of identifiers.
Are there split precincts? A single polling place may serve voters from two or more districts. The official results may report these splits separately. If the tabulator counted them together on one tape, you will need to understand how the splits are handled to avoid false discrepancies.
Are there notes or footnotes? Look for explanations of consolidation, corrections, adjustments, or reprocessing. These are often buried in fine print but can explain apparent discrepancies before you ever flag them.
AC’s approach to official results
AC has built automated pipelines to ingest official results at scale. The method varies by state and often by county: some states provide clean CSV downloads; others require scraping HTML tables; others require downloading and parsing PDF files. Where possible, AC downloads machine-readable data rather than transcribing from PDFs.
AC maintains per-jurisdiction knowledge about where results are published, in what format, and what quirks to expect — naming conventions, precinct numbering schemes, the presence or absence of separate voting-mode breakdowns. This knowledge accumulates over time: the first analysis of a new jurisdiction requires significant setup work to build the ingestion pipeline and understand the format; subsequent analyses of the same jurisdiction benefit from everything the first one taught.
9.6 — Step 4: Aggregation and Merging
This is the hard part. Not conceptually — the idea of matching one set of numbers to another is straightforward — but operationally. The poll tape shows votes counted by one tabulator at one precinct. The official results show votes aggregated and reported at whatever level and in whatever format the jurisdiction uses. Making these two datasets comparable requires understanding and replicating the jurisdiction’s aggregation and reporting logic. And that logic is often not fully documented.
This step is where most of the intellectual work of a comparison analysis happens, and where most of the non-obvious problems emerge. It is also where the most common source of false discrepancies originates: not errors in vote reporting, but errors in the analyst’s understanding of how vote reporting works in that particular jurisdiction. AC’s internal shorthand for this is “we did the comparison wrong” — and it is the single most common outcome of investigating an initial non-match.
Common aggregation challenges
Multiple tabulators per precinct. If a precinct has two scanners, the official precinct total is the sum of both tabulators’ tapes. If you have only one tape, your transcription will be lower than the official result by approximately the second tabulator’s contribution. This is expected, not a discrepancy. You need to determine which precincts have multiple tabulators (sometimes indicated on the tape, sometimes only discoverable through the official results or by contacting the elections office), and either obtain all tapes for those precincts or exclude them from comparison.
Split precincts. A single polling place may serve voters from two or more districts — congressional, state legislative, county commission, etc. A voter at the polling place receives a ballot specific to their district. The tabulator may count all ballots together and print a single tape with combined results, or it may separate results by district on the tape. Meanwhile, the official results may report these splits separately (e.g., “Precinct 42A” and “Precinct 42B” for different congressional districts). Understanding how the jurisdiction handles split precincts is essential. If the tape reports combined results and the official results report them split, you cannot make a comparison at the split level — you would need to sum the official split totals to compare against the tape.
Early voting and mail-in ballots. This is perhaps the most common source of confusion in a first analysis. The poll tape at the precinct typically captures only the votes counted by that tabulator — which usually means the election-day in-person votes processed at that location. Early voting ballots may be counted centrally, not at the precinct. Mail-in and absentee ballots are almost always counted centrally. Provisional ballots are adjudicated and counted after election day. If the official results combine all voting modes into a single precinct total, the official total will be higher than your tape transcription. This is expected. You need to either find official results that break out Election Day results separately, or exclude precincts where you cannot isolate the election-day component.
Provisional and late-arriving ballots. These are processed after election day and added to the official totals during the canvass. They will not appear on election-night poll tapes. If you are comparing against certified results (which include provisionals), the official total may be slightly higher than the tape total even for a single-tabulator, election-day-only precinct. The difference should be small and consistent with expected provisional ballot volumes.
Precinct identifier mismatches. The tape might print “PCT 042.” The official results might list the same precinct as “Precinct 42,” “42.0,” “0042,” “Eagle Creek,” or “EC-042.” Building a reliable crosswalk — a mapping between tape identifiers and official-results identifiers — is often the most tedious part of the merge. It requires careful comparison, sometimes manual inspection, and occasionally research into the jurisdiction’s precinct naming conventions. A wrong match in the crosswalk produces a false discrepancy; a missed match produces a missing comparison. Take your time with this.
Consolidated precincts. Some jurisdictions consolidate small precincts for reporting purposes — two or more precincts reported under a single unit. If your tape is from one of the component precincts, you need to sum all component tapes to match the consolidated official total. If you have only some of the component tapes, you cannot make a comparison.
Unexplained adjustments. Some official results include lines for adjustments, corrections, overvotes, undervotes, or “unresolved write-ins” that don’t correspond to anything on the poll tape. These must be understood before you treat a mismatch as a discrepancy. They may explain a small difference between your tape total and the official total.
Building the crosswalk
The crosswalk is the mapping between your transcription data and the official results. At minimum, it maps precinct identifiers between the two datasets. It may also need to map contest names (the tape says “US HOUSE DIST 11”; the official results say “U.S. Representative, 11th Congressional District”) and candidate names (the tape says “REP - SMITH, JOHN”; the official results say “John Smith (Republican)”).
For a small analysis, you can build the crosswalk manually — visually matching up identifiers between your spreadsheet and the official results. For a larger analysis, you may want to automate it: fuzzy string matching, regular expression cleaning, or a lookup table maintained across analyses of the same jurisdiction.
A good crosswalk is the foundation of a good analysis. Invest the time to get it right.
For small analyses: doing this by hand
If you are comparing a handful of precincts for one or two contests, you can do the aggregation and comparison manually in your spreadsheet. Build a table with one row per comparison: precinct identifier, contest name, transcribed vote total, official vote total, and a comparison column that flags matches and mismatches. Check each mismatch against the common aggregation challenges above before treating it as a real discrepancy.
For larger analyses: automation
For county-wide or state-wide analysis, manual comparison becomes impractical. Spreadsheet formulas (VLOOKUP, INDEX/MATCH) can automate the merge if identifiers are clean. Python — particularly the pandas library — is well-suited for larger datasets: load both data sources, build the crosswalk as a merge key, join the datasets, and compute comparison columns programmatically. AC uses automated Python pipelines for this step at scale.
The comparison table
The output of the aggregation and merge step is a comparison table: one row per comparison, with columns for precinct, contest, transcribed vote total, official vote total, and a match/mismatch flag. “Comparison” means you have a clean transcription, you have corresponding official results at the same granularity, and you have accounted for known aggregation factors (multiple tabulators, voting modes, split precincts, etc.).
This table is the core analytical product of a comparison analysis. Everything that follows — interpreting results, investigating discrepancies, asking the office about apparent discrepancies — operates on this table.
9.7 — Step 5: Interpreting Results
When everything matches
If every comparison in your table is a match — every transcribed vote total equals the corresponding official vote total — that is positive evidence that the vote reporting for those precincts and contests was accurate. It is evidence that the numbers printed on the tabulators’ tapes were faithfully transmitted through the reporting chain to the official certified results.
It is not a guarantee of election accuracy in the broader sense. The matched poll tape caveat (discussed in detail in Section 13.4) applies: a match between the tape and the official results confirms the reporting layer but cannot speak to the counting layer. If the tabulator miscounted ballots — misread marks, skipped ballots, experienced a calibration error — the tape would reflect the wrong count, the official results would reflect the same wrong count, and the comparison would show a match. Reporting accuracy and counting accuracy are different things, and a comparison analysis checks only the former.
But a high match rate across many precincts is meaningful evidence, and it accumulates. If your transcriptions are accurate and your merge logic is correct, lots of matches across many precincts and contests are good evidence that both your methodology and the jurisdiction’s reporting process are sound. This is a Bayesian inference: each additional match raises your confidence in the overall system, even though no individual match is proof. Conversely, if you find that everything matches, that should increase your confidence that the few remaining comparisons — including any mismatches — are meaningful.
Clean findings have practical value. They increase public confidence in the accuracy of the reported results. They validate the methodology and demonstrate that independent verification is feasible. They create a baseline for future elections — this county was clean in 2024; if discrepancies appear in 2026, that is new and noteworthy information. And they contribute to deterrence: the knowledge that independent verification is happening shifts incentives for anyone who might otherwise be careless with vote reporting.
When something doesn’t match
Do not panic. Most initial mismatches resolve to one of a handful of mundane explanations: a transcription error, a precinct identifier mismatch, a missing second tabulator tape, a voting-mode aggregation difference. A raw mismatch is an initial non-match — it is a signal that something needs to be investigated, not evidence that something went wrong.
The disposition of an initial non-match is determined by the investigation process described in the next section. Until that investigation is complete, maintain the neutral term: initial non-match. Not error. Not fraud. Not even “problem.” Just a mismatch that needs to be understood.
A note on AC’s prioritization
AC does not process every submission sequentially. The pipeline prioritizes producing the most high-impact comparisons: national and statewide races first, competitive races, large jurisdictions, clear recordings. AC may not transcribe every contest in every county, or every video from a given county, in the first pass. The goal is many high-priority comparisons, not exhaustive coverage in the initial analysis.
When AI-assisted transcription produces a match, that match is evidence — not proof — of correctness. It helps narrow the search space: if the AI reads a tape and the result matches the official total, the chance of a reporting error at that precinct for that contest is low (though not zero — the AI could have misread the tape in a way that coincidentally matched). Human reviewers can then focus their effort on the flagged mismatches, the ambiguous values, and the cases where the AI had low confidence.
Blurry recordings, blurry values, and ambiguous tapes may be skipped in initial passes and revisited later — or not revisited at all, if the analysis has already produced enough comparisons to support its conclusions. This prioritization is a reasonable allocation of limited resources toward maximum election transparency return. Every comparison has value; the goal is to produce as many as possible, starting with the most important.
9.8 — Step 6: Investigating Initial Non-Matches and Discrepancies
This section expands on the seven-check investigation process described in Section 8.9, providing practical guidance for anyone — America Counts or an independent analyst — triaging an initial non-match and investigating any discrepancy that survives. The process is the same regardless of who conducts it; the only difference is that AC has institutional experience and per-jurisdiction knowledge that an independent analyst may need to build from scratch.
The mindset
An initial non-match means your transcribed total doesn’t match the official total for a given precinct and contest. It does not mean that the official results are wrong. It does not mean that fraud occurred. It does not mean that you have found evidence of anything, yet. It means that two numbers that should be the same are not the same, and you need to find out why.
The goal of investigation is to determine what happened — not to confirm a suspicion. Approach every initial non-match with the assumption that there is a boring explanation, and look for it systematically. In AC’s experience, the boring explanation almost always exists: a transcription error, a merge mistake, a missing tabulator tape, an undocumented reporting convention. The seven checks below are designed to triage these initial non-matches — catching AC’s own errors and informing the investigation of a real discrepancy — before any heavier ask to the elections office.
If you work through all seven checks and cannot find an explanation, that is when the difference becomes a real discrepancy and the next phase of the process begins.
The seven checks
The seven-check investigation process is described in full in Section 8.9. Work through each check in order: (1) transcription error, (2) comparison logic error, (3) missing poll tapes, (4) duplicate tapes, (5) aggregation logic error, (6) undocumented vote reporting logic, (7) operational context. What follows here are practical tips that supplement the canonical descriptions in Section 8.9.
When a value is genuinely ambiguous — the ink is faint, the camera didn’t hold steady, the digit could plausibly be either of two values — flag it as unresolvable and exclude that comparison from your comparison table. Do not force a reading. An honest “I can’t read this” is more valuable than a guess that might produce either a false discrepancy or a false match.
Use discrepancy magnitude as a diagnostic. If the discrepancy is roughly half the official total, you’re probably missing a second tabulator’s tape (Check 3). If the discrepancy is roughly double, look for a duplicate submission (Check 4).
Watch for voting-mode mismatches (Check 5). If the official results combine all voting modes — Election Day, Early Voting, mail-in — and you have only the Election Day tape, the official total will be higher. Also check whether the jurisdiction consolidated precincts differently than you assumed, especially in low-turnout elections (primaries, special elections) where small precincts are more likely to be combined.
Jurisdiction quirks live in Check 6. Every jurisdiction has vote reporting conventions that aren’t documented anywhere public. Some include voided or adjudicated ballots in official totals but not on poll tapes. Some handle write-in candidates differently between the tape and the official tally. Some report absentee ballots under a separate “precinct” line (often called “ABS” or “MAIL” or “CENTRAL COUNT”) rather than distributing them back to physical precincts. When you hit Check 6, research: the elections office website, the canvass report, prior-year results (do they show the same pattern?), the state’s election procedures manual, and local news coverage. Keep your own per-jurisdiction knowledge notes — the knowledge you accumulate investigating one election’s discrepancies will accelerate the next.
For Check 7, cast a wide net for operational context. Search for news coverage, election office announcements, and social media posts from election day. Equipment issues — a tabulator jam that required re-scanning a batch of ballots — can cause the tape total to diverge from the official count in either direction.
If all seven checks pass
You have an apparent discrepancy that survives investigation. This is significant — not because it proves anything yet, but because you have done your due diligence and cannot explain the mismatch.
Document everything. Record your transcription, the official result, the discrepancy amount, which of the seven checks you performed, and what you found (or didn’t find) in each. This documentation is the foundation for everything that follows: the heavier ask to the elections office, any subsequent analysis, and the eventual published finding. An undocumented discrepancy is not credible; a thoroughly documented one is.
Proceed to the next step: a heavier ask to the elections office.
9.9 — Step 7: What to Do When You Find an Apparent Discrepancy
Before you contact anyone: additional due diligence
Before you involve the elections office, do everything you can to ensure that the discrepancy is real and that you haven’t missed an explanation. The seven-check process is thorough, but a few additional steps can increase your confidence.
Look at other submissions from the same precinct. If multiple AV volunteers recorded tapes at the same location, do the other recordings show the same numbers? If so, the transcription is almost certainly correct. If one recording shows different numbers, investigate whether it’s a different tape (a second tabulator, a zero tape vs. a results tape) or a different reading of the same tape.
Look at other precincts in the same county. Is the discrepancy isolated to one precinct, or is there a pattern? Multiple mismatches concentrated in the same geographic area, or following the same pattern (all in one direction, all of similar magnitude), is a stronger signal than a single isolated mismatch.
Look at prior elections in the same jurisdiction. Has this kind of discrepancy appeared before? Are there known reporting issues in this county? Prior analyses — by AC or by others — may have documented the same pattern.
Search for existing documentation. Look for post-election audits, canvass reports, press coverage, or public records requests that might already address or explain the discrepancy.
Get a second pair of eyes. If possible, have another analyst review your work independently. A fresh perspective catches errors that familiarity misses. Show them the discrepancy and your investigation trail, and let them try to resolve it.
The goal is to be as confident as possible that you haven’t missed something before you bring the discrepancy to the elections office. The credibility of the AV ecosystem — and of your own analysis — depends on bringing only thoroughly investigated discrepancies to the elections office.
Contacting the elections office
The appropriate first contact for an apparent discrepancy is the relevant county elections official: the Supervisor of Elections (in Florida), the Board of Elections (in North Carolina), the County Clerk, the Registrar of Voters, the Elections Director — the title varies by state. This is the office that administered the election and produced the official results. They have access to information that you do not: internal tabulation logs, chain-of-custody records, operational notes from election day, and the institutional knowledge of how their reporting system actually works.
Tone matters enormously. Elections officials are public servants who, in the vast majority of cases, are doing their best to administer elections accurately and fairly. They are also, in many jurisdictions, understaffed, under-resourced, and dealing with unprecedented levels of public distrust and even threats. The way you approach them determines whether you get a cooperative response or a defensive one.
The message should be collaborative, not accusatory. Specific about what you observed and what you’ve already checked. Asking for help understanding, not demanding answers. Respectful of the office’s expertise and workload. And brief — busy people respond better to concise messages.
Here is an example message:
Dear [Title and Name],
My name is [Name] and I am [briefly describe your role — “a volunteer with America Counts,” “an independent election transparency researcher,” etc.]. I am writing to ask for your help understanding a data question I’ve encountered while reviewing publicly available election records from the [Date] election in [County].
While comparing the precinct-level poll tape totals from [Precinct Name/Number] against the officially certified results, I noticed a difference in the [Contest Name] contest: the poll tape shows [X] votes for [Candidate], while the certified results show [Y] votes — a difference of [Z] votes.
I’ve checked several possible explanations on my end — transcription error, multiple tabulators at the precinct, differences in how voting modes are aggregated, and precinct identifier matching — but haven’t been able to account for the discrepancy.
I’d be grateful for any clarification your office can provide. It’s entirely possible that there’s a reporting convention or operational detail I’m not aware of that explains the difference, and I’d welcome the opportunity to understand it.
I’ve attached [or linked to] the relevant poll tape recording and my comparison data for your reference.
Thank you for your time, and for the work your office does administering elections in [County].
Sincerely, [Name] [Contact information]
This template can be adapted for your specific situation. The essential elements are: identify yourself, state what you found, state what you’ve already checked, ask for help, and express genuine respect for the office. Do not accuse. Do not speculate about causes. Do not use the word “fraud.” Present data and ask a question.
What to expect
America Counts’s experience shows that elections offices generally respond quickly and helpfully when contacted in this tone. In every case that AC has asked about a discrepancy to date, the office has provided a good explanation — a reporting convention that AC didn’t know about, a consolidation rule, an operational event, or another detail that resolved it to a match. The Henderson County correspondence is exemplary: AC contacted the Henderson County Board of Elections about a discrepancy at the Bat Cave precinct, the BOE responded the same day, confirmed the explanation (a second polling location opened due to hurricane damage), and provided the missing poll tape. The discrepancy was resolved collaboratively, and the exchange was documented publicly.
Allow reasonable time for a response. Elections offices are often small and understaffed, particularly in the weeks immediately after an election when they are still completing the canvass. At least two to four weeks for an initial response is reasonable. If you haven’t heard back, send a polite follow-up referencing your original message. AC’s policy is to wait at least a month, with at least one follow-up, before considering any further steps.
If the elections office resolves the discrepancy
This is the most common outcome, and it’s a sign that the system is working as it should. Document the explanation. Update your analysis to reflect the resolution — the discrepancy becomes a match, with the exchange recorded in its lineage. Thank the office — courtesy is important for maintaining cooperative relationships. And add the resolution to your jurisdiction knowledge notes so that future analysts don’t have to rediscover the same reporting convention.
A match with a non-trivial lineage is not a failure of your analysis — it is a success. You identified something that needed investigation, you investigated it thoroughly, you asked the office, and you received an explanation that resolved it. The investigation trail and the resolution together are valuable documentation of how vote reporting works in that jurisdiction.
If the elections office cannot or will not resolve the discrepancy
Several scenarios are possible. The office may provide an explanation that doesn’t actually account for the numbers — an explanation that explains part of the discrepancy but not all of it, or an explanation that doesn’t hold up when you check it against the data. In this case, engage constructively: explain what still doesn’t add up and ask if there is additional information. The office may have a valid point that you haven’t considered, or the explanation may genuinely be insufficient.
The office may not respond at all, despite follow-up attempts. Non-response is frustrating but not uncommon, particularly for offices that are under-resourced or that have adopted a policy of not engaging with external inquiries.
The office may dispute your methodology or your data. Engage on the merits — they may be right about something — but document the exchange regardless.
In any of these scenarios, you now have an apparent discrepancy with a documented investigation trail and a documented ask history. This is the point at which the question shifts from “what happened?” to “what should happen next?”
Escalation pathways
The nature of the discrepancy, its magnitude, the contest it affects, and the response from the elections office all inform what comes next. The legal pathways paper — Legal Pathways for Actual Vote Evidence, prepared as a companion document for this manual — provides a comprehensive analysis of the legal and procedural options available when AV evidence reveals an apparent discrepancy that no cooperative ask has resolved. What follows is a summary of the six pathways it identifies, in approximate order of practical accessibility. Readers who need operational detail — evidentiary standards, standing requirements, procedural timelines, likely objections and responses, and a worked example using Florida law — should consult the full paper.
1. Regulatory complaints
A regulatory complaint is a formal submission to an administrative body with jurisdiction over election administration. This is the lowest-bar escalation pathway: no formal standing requirement (anyone with credible information can file), no formal evidentiary bar (administrative reasonableness, not the Rules of Evidence), and flexible timelines (no tight statutory deadline).
Potential recipients include: the state board of elections or elections commission, the Secretary of State, the state Attorney General, the federal Election Assistance Commission, and the U.S. Department of Justice Civil Rights Division (Voting Section) if federal elections are affected or if the discrepancy disproportionately affects protected groups.
Regulatory action can produce real and lasting results. The Monmouth County, New Jersey case demonstrates what this pathway can achieve: the New Jersey Attorney General’s office launched an independent investigation, retained outside counsel, published detailed findings, obtained a court-ordered recount, and secured a corrected certification — all triggered by the discovery of a vote-counting irregularity. AC’s evidence package, with its documented methodology and transparent investigation trail, is well-suited to regulatory complaint processes.
2. Media and public transparency
The AV archive is public. Your analysis can be too. Sharing your findings — your data, your methodology, your investigation trail, and the elections office response (or non-response) — through public channels serves the transparency mission directly. Investigative journalists have been essential in surfacing election integrity issues: the Shelby County story, the Butler County story, and the Prince William County story all benefited from press attention that produced public accountability.
AC publishes its findings as analysis reports (described in Section 12). Independent analysts can do the same — through academic publications, blog posts, civic organization reports, or direct outreach to journalists covering election administration.
3. Supporting candidate action
If the discrepancy affects a specific contest, the affected candidate or their campaign may have legal standing to file a state election contest — a statutory procedure available in every state by which a candidate challenges the certified result. State election contests have tight statutory deadlines, typically five to ten days after certification. If you believe your findings may support a contest petition, timing matters: the candidate needs to know before the deadline expires.
AV evidence — the poll tape recording, the transcription, the comparison analysis, and the investigation trail — can serve as the factual basis for a contest petition. The legal pathways paper analyzes the evidentiary standards for this use in detail.
4. Public records litigation
If the elections office is uncooperative — refusing to provide the information needed to resolve the discrepancy, or refusing to respond at all — public records laws (state FOIA equivalents) provide a legal mechanism to compel disclosure. AV evidence can support a public records request by demonstrating a specific, documented basis for seeking the underlying election records.
5. Federal civil rights litigation
If the discrepancy disproportionately affects voters based on race, ethnicity, or other protected characteristics, federal civil rights statutes may apply. This is a higher-bar process that requires legal counsel and, typically, a broader factual record than a single discrepancy. AV evidence can serve as one component of the evidence in a civil rights case — demonstrating that specific reporting errors affected specific communities.
6. Criminal referral
If the investigation suggests intentional misconduct — not just error, but deliberate manipulation of vote reporting — the matter can be referred to the relevant prosecuting authority: the state attorney general, the county prosecutor, or the federal Department of Justice. AV evidence alone does not prove intent, identify a perpetrator, or establish the elements of a crime. It documents a discrepancy. Criminal investigation is the highest-bar pathway and is beyond what AC or any independent analyst pursues directly. But a documented, thoroughly investigated discrepancy that cannot be explained provides the factual basis that a prosecutor would need to determine whether further investigation is warranted.
9.10 — A Note on the Nature of This Work
This is hard
Vote reporting analysis of this kind is genuinely difficult and often tedious. It requires patience, attention to detail, comfort with ambiguity, and a willingness to spend hours researching the specific reporting conventions of a jurisdiction you’ve never previously studied. The reason this work hasn’t been systematically done before is not that nobody thought of it — it’s that the operational challenges are real.
Vote reporting in the United States is not standardized. There is no national format for official election results. There is no standard precinct numbering system. There is no universal convention for how voting modes are categorized, how split precincts are handled, how adjustments are applied, or how results flow from precinct tabulators to county election management systems to state reporting databases to the public record. Every state has its own rules. Many counties have their own conventions within those rules. And much of the detailed logic is undocumented.
Every time AC conducts a comparison analysis in a new jurisdiction, it encounters things in the vote reporting that are not obvious and have to be researched, hypothesized about, tested, and checked. Prior experience helps — the patterns repeat, the categories of problems are finite, and institutional knowledge accumulates — but there is always something new. A comparison analyst should expect to spend a significant fraction of their time not on the comparison itself, but on understanding how the jurisdiction reports its results.
Incomplete information is the norm
In an ideal world, every jurisdiction would publish complete technical documentation of its vote reporting logic: how precincts map to reporting units, how voting modes are aggregated, how split precincts are handled, how adjustments and corrections are applied, how results flow from the tabulator to the official certified record. In practice, this documentation is rarely available in complete form. Some jurisdictions publish procedure manuals. Some publish canvass reports with useful detail. Some publish nothing beyond the results themselves.
This means that much of the analytical work involves educated inference. You study the data, look for patterns, form hypotheses about the reporting logic, test those hypotheses against additional data points, and arrive at the best explanation available. When your hypothesized logic produces matches across most precincts, that’s strong evidence that your logic is correct — and the few mismatches are worth investigating. When your logic produces mismatches everywhere, the problem is almost certainly your logic, not the reporting.
There is a Bayesian dimension to this. If you follow reasonable steps for the comparison analysis and you get a high match rate — lots of precincts matching, across multiple contests — that is good evidence that both your methodology and the jurisdiction’s reporting are sound. It’s not a guarantee. The matches could be coincidental, your logic could be wrong in a way that happens to produce matches, or both you and the jurisdiction could be making the same error. But it raises your confidence in both your work and the official results — probabilistically, not certainly. Conversely, a cluster of mismatches in an area where everything else matches is a stronger signal than a single mismatch in a dataset full of mismatches.
The research mindset
Approach this work with a research mindset, not an adversarial one. You are producing independent evidence about the accuracy of vote reporting — not specifically looking for discrepancies because you believe they must exist. But maintain an auditor’s rigor: for every comparison possible in your data, make sure it’s valid, and if it’s not, investigate and — when warranted — ask the office.
This distinction matters for credibility. An analyst who approaches every mismatch as evidence of wrongdoing will produce findings that nobody trusts — because the findings will be contaminated by confirmation bias. An analyst who approaches every mismatch as a puzzle to be solved, and who follows the evidence wherever it leads, will produce findings that are credible to elections officials, journalists, courts, and the public.
Take reasonable steps to drive toward a reasonable-looking answer. If you can’t get to certainty, document what you know and what you don’t know. Uncertainty, honestly documented, is more valuable than false precision. “We found a 47-vote discrepancy that we cannot explain after checking the following seven things” is a credible finding. “We proved that 47 votes were stolen” is not — because this analysis cannot prove theft, and claiming otherwise discredits the work.
A note on documentation
Better public documentation of vote reporting logic from elections offices would make this work dramatically easier and more reliable. Sometimes you can find what you need: canvass reports, procedure manuals, state election law, elections office FAQs. Often you can’t, and you proceed with the best information available.
AC’s recommendations to elections offices (Section 15) include specific requests for improved documentation and machine-readable results — published in standardized formats, with precinct-level granularity, with voting modes broken out separately, with clear precinct identifier schemes. These recommendations are not criticism of current officials. They are a recognition that transparency works better when the underlying data is accessible, and that the current patchwork of formats and conventions makes independent verification harder than it needs to be.
9.11 — How America Counts Does This at Scale
This section describes AC’s institutional pipeline for readers who want to understand the high-throughput process. Independent analysts who are conducting their own analysis can skip this section — the logical steps are the same as described above; AC simply automates and scales them.
The high-throughput pipeline
AC’s goal is to maximize the election transparency return on every poll tape submission — to produce as many comparisons as possible, as quickly as possible, with appropriate quality controls. The pipeline is designed to handle hundreds or thousands of submissions across multiple states and dozens of counties in a single election cycle.
The pipeline automates as much as possible while maintaining human oversight at critical decision points. The automation handles the repetitive, high-volume work: processing videos, reading values, matching identifiers, computing comparisons. The human oversight handles the judgment calls: resolving ambiguous transcriptions, investigating apparent discrepancies, interpreting undocumented reporting logic, and deciding when a discrepancy has been sufficiently investigated.
Transcription at scale
AC uses AI-assisted transcription to process poll tape videos. The AI — optical character recognition applied to video frames — reads the tape and produces an initial transcription for each contest and candidate. This first pass is fast and handles the majority of clear, well-recorded tapes accurately.
Every AI-generated transcription is subject to human review. The AI’s output is checked against the source video by a human reviewer who confirms or corrects each value. When the AI reads a value with low confidence, when the video quality is poor, or when the tape layout is unusual (non-standard formatting, unusual fonts, smudged ink), the human reviewer makes the final determination.
AC prioritizes high-quality recordings over marginal ones. Clear videos of important contests are processed first. Blurry recordings, partially obscured tapes, and submissions with significant quality issues may be deferred or skipped in the initial pass and revisited later if resources allow. The principle is simple: maximize the number of high-confidence comparisons per unit of effort.
Official results ingestion
AC has built automated scripts to download, scrape, or ingest official results from the jurisdictions it analyzes. The method depends entirely on the format: some states provide clean CSV downloads from a centralized portal; others require scraping HTML tables from county websites; others require downloading and parsing PDF documents; a few require manual entry from scanned images.
AC maintains these ingestion scripts on a per-state and sometimes per-county basis. When a new jurisdiction is encountered for the first time, building the ingestion pipeline is part of the analysis setup. Once built, the pipeline can be reused for future elections in the same jurisdiction — often with minor adjustments for format changes.
Comparison execution
Automated Python pipelines handle the aggregation and merge at scale. The pipeline builds the precinct crosswalk (mapping tape identifiers to official-results identifiers), joins the transcription data to the official results, computes comparison columns, and flags mismatches. The output is the comparison table: one row per comparison, with match/mismatch classification, discrepancy magnitude, and metadata about the source recordings and official results.
Human review and investigation
The comparison table is reviewed by hand. Matches are spot-checked against source videos to verify that the AI transcription was correct. Mismatches enter the seven-check investigation process described in Section VIII above. In AC’s experience, the vast majority of apparent discrepancies resolve through one of the seven checks — most commonly transcription error (Check 1), comparison logic error (Check 2), or undocumented vote reporting logic (Check 6).
The investigation step is where AC’s accumulated per-jurisdiction knowledge pays off. Each match with non-trivial lineage teaches AC something about how that jurisdiction reports its results. These lessons are captured in per-jurisdiction knowledge notes that make subsequent analyses faster and more accurate. The first analysis of a new jurisdiction is always the hardest; subsequent ones benefit from everything the first taught.
Prioritization logic
AC’s pipeline prioritizes the most important races: national contests (presidential, U.S. Senate, U.S. House) first, then statewide contests (governor, attorney general, secretary of state), then competitive local races. Within each tier, the pipeline may focus on the largest jurisdictions, the jurisdictions with the most AV submissions, or the jurisdictions where prior analysis has raised questions.
When the AI-assisted transcription produces a match, that is treated as evidence (not certainty) of correctness. It allows human reviewers to focus their limited time on the flagged mismatches and ambiguous cases rather than reviewing every single comparison from scratch. This is how the pipeline scales: the AI handles the high-volume clear cases; humans handle the difficult ones.
The goal is many high-priority, high-impact comparisons rather than exhaustive sequential processing of every submission. A comparison of the presidential race in a major county is more impactful than a comparison of a school board race in a small township — both are valid, but when resources are limited, the higher-impact comparison comes first.
Heavier asks and escalation pathways
Any apparent discrepancy that survives the seven-check process is brought to the relevant elections office through a heavier ask, following the process described in Step 7 above. AC documents all asks and their outcomes. The Henderson County/Bat Cave example — same-day response, collaborative resolution, public documentation — is the model AC aims for.
In cases where the elections office cannot or will not resolve a discrepancy, AC turns to the escalation pathways described in Step 7, with the full legal analysis available in the companion legal pathways paper.
Reporting
Findings are compiled and published as analysis reports. Section 12 of this manual describes the two report formats (static analysis reports and rolling operational logs), the vocabulary discipline that governs how discrepancies are characterized, and the standards for presenting both discrepancy findings and clean results. Whether or not discrepancies are found, the analysis is published — because clean results are themselves evidence of reporting accuracy and contribute to the transparency mission.
9.12 — For the Record: What This Analysis Can and Cannot Show
A comparison analysis is a powerful tool, but it is a specific tool with specific capabilities and specific limitations. Intellectual honesty about what it can and cannot do is essential for the credibility of every analysis, every analyst, and the AV ecosystem as a whole.
What it can show
A comparison analysis can show whether independently-captured poll tape totals match officially reported results at the precinct level. It can quantify the scope of any discrepancy: which precincts are affected, which contests, and how many votes. It can document whether an apparent discrepancy can be explained by known factors (transcription error, aggregation differences, undocumented reporting logic, operational events) or whether it remains unresolved after thorough investigation. And it can calculate the overall match rate across all comparisons — a measure of reporting accuracy for the covered precincts and contests.
A high match rate across many precincts is meaningful positive evidence: evidence that the vote reporting in those precincts, for those contests, was accurate. An apparent discrepancy that no ask could resolve is meaningful negative evidence: evidence that something in the reporting chain produced a different number than what the tabulator printed. Both findings contribute to the public record.
What it cannot show
A comparison analysis cannot show whether machines counted ballots correctly in the first place. This is the matched poll tape caveat, discussed in detail in Section 13.4: if a tabulator miscounts ballots, the poll tape will reflect the wrong count, the official results will reflect the same wrong count, and the comparison will show a match. A match confirms the reporting layer — not the counting layer.
A comparison analysis cannot identify who is responsible for a discrepancy, or whether a discrepancy was intentional or accidental. It documents what the numbers are; it does not explain why they differ. That determination requires additional investigation, typically by the elections office, a regulatory body, or a legal proceeding.
A comparison analysis cannot determine whether a discrepancy affected the outcome of an election. That requires comparison of the discrepancy magnitude against the margin of victory — a statistical and legal question, not a transcription question.
And a comparison analysis cannot guarantee that the absence of discrepancies means the election was conducted correctly. It means the reporting layer checks out for the covered precincts. The counting layer is not tested. Precincts without AV coverage are not tested. And the analysis is only as good as the transcription, the merge logic, and the official results data that went into it.
The value of doing it anyway
Even with these limitations, a comparison analysis produces something that currently does not exist for most American elections: independent evidence about the accuracy of the reported results. That evidence is valuable whether or not discrepancies are found. Clean results increase public confidence. Discrepancies, when found and resolved, improve understanding of the reporting process. Discrepancies that cannot be resolved demand further investigation.
The evidence is independent — captured by citizens, held outside government custody, permanently archived. The methodology is transparent — documented in this manual, reproducible by anyone, and open to scrutiny. The findings are published — available for public inspection, journalistic investigation, academic research, and legal proceedings.
This is what election transparency looks like in practice. It is not perfect. It is not comprehensive. It is, right now, the only systematic independent check on the reporting layer of American elections. And it is something that any citizen with a smartphone, a spreadsheet, and the willingness to do careful work can contribute to.