12 — Findings and Reports

This section is about the output side of Actual Vote: how AC presents its findings, what formats reports take, how discrepancies are characterized, and how findings move into the world. It builds on the methodology described in Section 8 but takes a broader view, focusing on the relationship between the analysis and its readers rather than on the mechanics of producing the analysis.

There are six subsections. Section 12.1 describes the anatomy of an Actual Vote finding — what it consists of, what principles govern its presentation, and what scope it carries. Section 12.2 describes the two report formats currently in use. Section 12.3 develops the vocabulary for characterizing discrepancies — one of the most important aspects of AC’s public voice. Section 12.4 covers publication, communication, and the Analysis Window — where reports go, who sees them, how timing affects presentation, and what happens after publication. Section 12.5 addresses the interpretation of “clean” results, which is where AC’s most important findings to date have fallen. Section 12.6 sets out how AC reasons under incomplete documentation — the epistemic stance behind every number in a report, and the discipline that keeps inference honest.

A reader who wants to understand what an AC report looks like should read Section 12.1 and Section 12.2. A reader who wants to understand how AC talks about problems should read Section 12.3. A reader evaluating what an AC finding means should read Section 12.5.

12.1 — What an Actual Vote Finding Consists Of

A finding is a comparison plus its disposition

The atomic unit of an Actual Vote finding is a single comparison from the comparison dataset, with its disposition. Each comparison says: for this County × Precinct × Voting Method × Contest × Choice combination, the aggregated transcribed values were X, the official value was Y, and the result was a match, a likely attributable discrepancy, an apparent discrepancy, escalatable, escalated, or a true discrepancy. The finding for that comparison is the combination of the observation (X and Y) and the disposition (the result of investigation).

A whole analysis has many comparisons. The headline finding of an analysis is typically a summary of the comparison results: AC performed N comparisons covering M votes; of these, K were exact matches, J were likely attributable discrepancies, L were apparent discrepancies, and after investigation, all resolved to matches. The NC 2024 Primary Report is the canonical example of this kind of summary finding.

The headline finding rolls up the underlying comparisons, but the underlying comparisons are still the substance. A reader who wants to evaluate the finding can go to the comparison dataset and inspect every individual comparison. A reader who wants to challenge the finding has to challenge specific underlying comparisons, not just the rollup.

Observation vs. interpretation

The most important distinction in finding presentation is between observation and interpretation.

An observation is a fact about what AC measured. “The poll tape from precinct 03A in Wake County, recorded by submission 1524, shows 247 votes for Joe Biden in Election Day. The official results from the NCSBE show 247 votes for Joe Biden in Wake County precinct 03A in Election Day. The values match exactly.” This is an observation. It does not depend on AC’s interpretation; it depends on AC having read the poll tape correctly and downloaded the official results correctly.

An interpretation is what AC says the observation means. “Therefore, the vote reporting for this comparison is accurate.” That is an interpretation. It draws a conclusion from the observation. In most cases the interpretation is uncontroversial — an exact match means the reporting was accurate — but the interpretive step is real, and AC’s job is to be clear about which is which.

The distinction becomes more important when the observation is ambiguous. “The poll tape from precinct 04B shows 312 votes; the official results show 311 votes; we do not know why they differ; we asked the elections office and did not get a response.” That is an observation. Whether it supports the interpretation that “vote reporting was inaccurate by one vote” is a separate question. The observation does not depend on AC’s willingness to interpret; the interpretation does.

Findings are claims, not actions

A separate principle, drawn from AC’s published reasoning: an Actual Vote finding is a claim about vote reporting, not an action in the world. AC publishes findings; AC does not take action based on those findings. Action is the work of others — courts, elections offices, candidates, journalists, citizens — based on the evidence AC has provided.

This is a deliberate division of labor and a meaningful contributor to AC’s legitimacy. AC is the evidence-gathering organization; downstream actors are the action-taking ones. The implications for finding presentation are significant. Findings should be calibrated to support downstream evaluation, not to drive immediate action. A finding presented as “vote reporting was accurate” is more useful to downstream actors than one presented as “no further action is needed.” Findings should leave room for the reader to disagree — a finding presented as “AC observed exact matches in N of N comparisons” lets a skeptical reader verify the claim. Findings should be clear about what AC has and has not done — if AC did not ask the elections office about a particular apparent discrepancy because the certification deadline had passed, the finding should say that. And findings should not recommend specific actions to specific audiences; AC can describe what kinds of actions are possible, but the choice of action belongs to the actor.

Findings are not endorsements

Actual Vote findings are not endorsements of any candidate, party, position, or outcome. Findings describe vote reporting, not winning. Findings do not take sides on contested races. Findings do not validate or invalidate specific narratives — if a particular political narrative claims that vote reporting was fraudulent, an AC finding of “we found no apparent discrepancies” provides evidence about what AC measured, which the reader can integrate with whatever else they know. The non-partisan framing is part of why AC’s findings can be trusted by readers across the political spectrum. The moment AC starts making partisan endorsements, the findings become harder to trust regardless of their accuracy.

Findings are bounded in scope

Every Actual Vote finding has a scope, and the scope matters for what the finding does and does not claim.

Scope by election: a finding from the NC 2024 primary analysis is a finding about that election. It does not extend to the 2024 General Election in NC, to other elections in other states, or to elections in years AC did not analyze. Scope by jurisdiction: a finding covering seven NC counties does not extend to the other NC counties that were not analyzed. Scope by voting method: a finding may cover Election Day vote reporting but not Early Voting, depending on which voting methods were in scope. Scope by contest: even within a single election in a single jurisdiction, the analysis may focus on specific contests without auditing every contest on the ballot. Scope by what the poll tapes show: poll tapes show machine-tabulated totals, not the underlying ballots, the voter check-in data, the ballot order tracking, or many other parts of the election system. The matched poll tape caveat (Section 13.4) is the explicit acknowledgment of this scope limitation.

The scope should be made clear in every finding. A reader who does not understand the scope can over-extrapolate the finding to claims AC never made.

12.2 — Report Format and Structure

Two report formats

America Counts currently uses two report formats, each suited to different circumstances.

The static analysis report is a single document published once an analysis is complete. It describes what AC did and what AC found in a structured, definitive format. The NC 2024 Primary Report is the canonical example of a static report. Static reports are appropriate when the analysis has a clear endpoint, the findings are stable, and the audience expects a citable, archival document. The static format follows a structured sequence. The title and front matter include the Analysis Status (Open, Closed, or Pending) and the Analysis Window (before-certification or after-certification, with the relevant certification date). An overview section describes the methodology in plain language accessible to a non-technical reader. A scope section makes the analysis boundaries explicit: which state, which counties, which election, which voting methods, which contests, which official results source. The headline findings present the core numbers: total transcribable submissions, total transcribed values, number with apparent discrepancies, number with surviving discrepancies after investigation, and total votes audited. An impact statement articulates what the analysis demonstrates — using precise language like “provides evidence that” rather than “proves that,” “the N votes that it analyzed” rather than “all the votes,” and “reported correctly” rather than “the election was fair.” A limitations section makes the caveats explicit: the matched poll tape caveat, the coverage caveat, the throughput caveat, and the no-guarantees caveat from the User Agreement. The detailed findings section breaks results down by county, precinct, or contest, and documents any notable investigations. Acknowledgments credit volunteers, partner organizations, cooperating elections offices, and donors. Appendices provide supporting material for replicability: methodology detail, comparison code or pseudocode, source data references, submission references in the Actual Vote Video Archive, and a glossary. The static format is designed to be self-contained: a reader should be able to understand the analysis without reference to other documents, though cross-references to this manual and to the Actual Vote Video Archive are provided for readers who want deeper detail.

The rolling operational log is a live-updated page maintained during an active analysis. It captures findings, observations, and operational notes as they occur, in reverse-chronological order, with timestamps. The 2024 General Election News Page is the canonical example. The rolling format is appropriate when the analysis is ongoing, when new information arrives over days or weeks, and when the audience wants to follow the analysis in real time rather than wait for a final report. The rolling format naturally accommodates the straddle between before-certification and after-certification windows: entries made before certification are distinct from those made after, and the reader can see the evolution of the analysis across the transition. The rolling format trades the structural clarity of the static format for timeliness and transparency.

The NC 2024 Primary Report as exemplar

The NC 2024 Primary Report established the template for the static format. Its structure has become the recommended pattern for future static reports.

The report covers a bounded analysis: one election (the March 5, 2024 primary) in one state (North Carolina), spanning seven counties, with 52 transcribable submissions producing comparisons covering 321,502 individual votes. The headline finding is zero true discrepancies — every comparison either matched exactly or was resolved through investigation. The report was published in the before-certification window, giving it operational relevance as well as archival value.

Key features of the report that future reports should emulate: the front matter clearly states the Analysis Status and the Analysis Window. The methodology is described in plain language accessible to a non-technical reader. The scope is explicitly bounded — the report is careful to say that it covers the analyzed precincts, not the entire state. The findings are presented with the observation-interpretation distinction clearly maintained. And the matched poll tape caveat appears in the conclusions, reminding the reader what the analysis does and does not claim.

The 2024 General Election News Page as exemplar

The 2024 General Election News Page established the template for the rolling format. It covered a multi-state operation (North Carolina, Georgia, Florida) during the November 2024 general election and its aftermath. Entries documented findings as they were produced, operational challenges as they arose (app store rejections, weather disruptions, equipment failures), and investigation updates as apparent discrepancies were resolved.

The rolling format is less structured than the static format, which makes it more difficult to cite and more difficult for a new reader to parse. But it captures information that the static format cannot: the chronology of the analysis, the evolution of the investigator’s understanding, the operational realities that shaped the work. For audiences following the analysis in real time — journalists covering the election, partner organizations coordinating with AC, elections offices responding to inquiries — the rolling format provides the timeliness that the static format lacks.

Structure of a rolling log

A rolling log does not have a fixed section structure the way a static report does. Instead, it has a header with metadata and a sequence of dated entries.

The header at the top of the page typically includes: the title (the election or analysis name), a last-update timestamp (when the page was last edited), the current scope (which states, counties, and elections the page is covering), the current status (active, in transition to static report, or closed), a brief overview orienting the reader to what the page is and what they can find there, and cross-references to relevant background material (the methodology, the app page, previous election cycle news pages if relevant).

The body is a sequence of dated entries. Each entry is one or more paragraphs documenting what happened that day or in that period. The format is informal and journal-like. Common entry types include submission status updates (“Received N submissions from [team / county] today”), pipeline progress notes (“Vetting backlog at N submissions; transcription has begun for [county]”), specific incident documentation (“User [team] reports trouble with the Android app; investigating”), apparent discrepancy investigations (“Investigating apparent discrepancy in [precinct] for [contest]; current hypothesis is [reason]”), ask status (“Sent a heavier ask to [county BOE] regarding [issue]; awaiting response”), resolution notes (“[County BOE] responded; the apparent discrepancy was caused by [explanation]; reclassifying as a match”), operational notes (“Google Play Store rejected the latest Android build; users should use [fallback path]”), and methodological notes (“We’ve revised the comparison logic for [jurisdiction] to handle [edge case]; rerunning the analysis”). Entries range from a single sentence to several paragraphs depending on what happened.

Some rolling logs include aggregated issue sections that organize issues by status: an open-issues section listing apparent discrepancies still under investigation, pending asks to elections offices, and unresolved operational issues; and a resolved-issues section listing apparent discrepancies that have since resolved to matches, with their explanations. These aggregated sections give the reader a quick way to see the current state of the analysis without reading every chronological entry. They should be updated as issues move between open and closed.

A rolling log may also include running totals of the analysis state: total submissions received, total submissions vetted, total comparisons performed, total apparent discrepancies, total asks sent, total responses received. These are not the final numbers — the analysis is still ongoing — but they show the scale of the work and help readers track progress.

Pitfalls of the rolling format and how to mitigate them

The rolling format has specific risks that the static format does not. Stale content is the most common: entries written days or weeks ago may no longer be accurate, and the rolling log needs active maintenance to avoid leaving outdated claims in place. The mitigation is to update entries when they become stale, with each revised entry clearly noting that it has been updated and what changed. Inconsistent tone develops when entries written under time pressure during election week use different language and voice than entries written later during the analysis phase; maintaining a header summary at the top of the page that always reflects the current state helps orient readers regardless of when individual entries were written. Premature claims are entries that state a finding that later investigation proves wrong — the original entry stays in the page record even if the finding is later revised, and future readers may see the original without seeing the revision; AC mitigates this by clearly marking revisions and linking back to original entries. Overwhelming length makes long rolling logs hard to navigate; a table of contents at the top helps readers find specific entries. Missing context affects new readers arriving in the middle of the analysis who lack the background to understand entries; cross-referencing background material (this manual, the methodology page, AC’s institutional content) ensures new readers have somewhere to go for orientation.

The relationship between rolling logs and static reports

A rolling log is not a substitute for a static report. The rolling log is a process document — it documents what is happening, when, and why, but it is vulnerable to the pitfalls described above. The static report is a product document — it documents what was found, definitively, in a stable form that is easier to cite, easier to navigate, and more polished. For analyses that benefit from both, the typical lifecycle is: the analysis begins with a rolling log during the active phase; as findings stabilize, the rolling log accumulates them; at the end of the active phase, AC drafts a static report based on the rolling log content; the static report is published as the canonical summary; the rolling log may stay live as a reference or may be superseded by the static report. The 2024 General Election analyses are an example of work that started as a rolling log; as of the writing of this manual, the static report for the 2024 General Election has not yet been published, and the rolling log on the news page remains the most complete published account of that analysis.

12.3 — Characterizing Discrepancies

Vocabulary discipline

How AC talks about differences between poll tape values and official results is one of the most important aspects of its public voice. The terminology is not arbitrary — each term marks a specific place in a comparison’s life and a specific level of certainty about the nature of the difference. The complete map of how one comparison moves from first pass to a final outcome is collected in the glossary under comparison outcomes; this section explains the vocabulary and the judgments behind it.

The single most important distinction is between a non-match that is an artifact of our own process and a real difference in the reporting record. The first is never called a discrepancy.

From comparison to outcome

Every comparison — the mapping of one or more poll-tape transcriptions to a single official total — is, on its first pass, either an exact match (the numbers agree) or an initial non-match (they do not). An initial non-match is not yet a discrepancy; it is the raw funnel that review then triages. Most initial non-matches are AC’s own doing — a transcription slip, a matching-logic error, a mis-mapping of contest, choice, voting method, or precinct — and are corrected to a match. These are an expected, routine part of the work, and AC does not report them. (AC may note a genuinely interesting reporting quirk it had to model — an unusual aggregation, say — but not the routine corrections themselves.)

What survives review as a real difference in the reporting record is a discrepancy. Every discrepancy is described along three independent axes: its magnitude, whether it is contest-flipping, and its explanatory state. The axes are orthogonal — any combination can occur — and none of them bears on whether a discrepancy is worth resolving. All discrepancies are worth resolving; the axes calibrate urgency and effort, and shape how the discrepancy is reported.

The whole lifecycle — every path a single comparison can take, from first pass to a final outcome — is the map below. Click it to expand to full size.

flowchart TD
  A["One comparison"] --> B{"First pass"}
  B -->|equal| M["MATCH"]
  B -->|not equal| INM["Initial non-match<br/>(any reason)"]
  INM --> R{"AC reviews —<br/>form a hypothesis if we can"}
  R -->|"our own error:<br/>transcription / matching / mapping"| M
  R -->|"benign reporting-logic hypothesis"| LA["Likely attributable<br/>discrepancy"]
  R -->|"no benign hypothesis;<br/>could be error or fraud"| AP["Apparent<br/>discrepancy"]
  LA -->|"AC check confirms<br/>benign explanation"| M
  LA -->|"plausible explanation accepted;<br/>non-contest-flipping & benign magnitude"| MPE["MATCH<br/>(plausible explanation)"]
  LA --> ASKA{"Lighter ask<br/>to the SOE"}
  ASKA -->|"SOE confirms<br/>benign explanation"| M
  ASKA -->|"SOE offers plausible explanation;<br/>non-contest-flipping"| MPE
  ASKA -->|"no satisfactory<br/>resolution"| AP
  AP -->|"AC finds<br/>satisfactory resolution"| M
  AP --> ASKB{"Heavier ask<br/>to the SOE"}
  ASKB -->|"SOE gives<br/>satisfactory resolution"| M
  ASKB -->|"SOE acknowledges<br/>the error"| TD["TRUE DISCREPANCY<br/>(in this version of<br/>the official results)"]
  ASKB -->|"no satisfactory<br/>resolution"| ESCB["Escalatable<br/>discrepancy"]
  ESCB -->|"handed to a<br/>court-capable party"| ESCD{"Escalated discrepancy<br/>(court case underway)"}
  ESCB -->|"court case fails<br/>to materialize"| UNRES["Unresolved — no court case<br/>(discrepancy stands)"]
  ESCD -->|"court finds<br/>AV evidence correct"| TD
  ESCD -->|"official results<br/>ruled correct"| ORC["Official results ruled correct<br/>(poll-tape evidence fails to persuade court)"]
  ESCD -->|"something else<br/>happens"| UNC["Uncharted<br/>territory"]
  TD -->|"official results<br/>restated"| MNEW["MATCH<br/>(new official results)"]
  classDef good fill:#e7f5ec,stroke:#2e7d4f,stroke-width:2px,color:#114433;
  classDef bad fill:#fdecea,stroke:#b3261e,stroke-width:2px,color:#441111;
  classDef neutral fill:#eef1f6,stroke:#5a6577,stroke-width:1.5px,color:#222222;
  classDef unknown fill:#f4f1e9,stroke:#9a8c66,stroke-width:1.5px,color:#3a3320,stroke-dasharray:5 3;
  class M good;
  class MPE good;
  class MNEW good;
  class TD bad;
  class ORC neutral;
  class UNRES neutral;
  class UNC unknown;

Magnitude — absolute and relative

Magnitude is the size of a discrepancy, and AC reports it as two numbers rather than a label. The absolute magnitude is the number of votes in question — the difference between the transcribed and official values. The relative magnitude is that difference as a share of a stated denominator, normally the contest total at the comparison grain. Both matter, in different ways: ten votes is a large relative magnitude in a twenty-vote contest and a small one in a two-thousand-vote contest.

AC deliberately reports the figures rather than sorting discrepancies into “small” and “large” buckets. A fixed low/high label would be a judgment dressed up as a measurement; the numbers keep every characterization falsifiable. Magnitude governs the urgency and effort AC brings to a resolution — proportionally — but it never determines whether a discrepancy is worth pursuing. Even a one-vote discrepancy, once corrected, improves the accuracy of the record, deters larger future errors, and signals that citizen auditors are watching. (Magnitude replaces the older “material / immaterial” language, which implied that small discrepancies do not matter — a value judgment AC does not make.)

Contest-flipping

A discrepancy is contest-flipping when its resolution could change the true winner of the contest — that is, when the contest’s unresolved difference, measured against the margin of victory and accounting for direction, is large enough to decide the outcome. The complement, non-contest-flipping, is the default, and AC assumes it unless stated otherwise. This axis is orthogonal to magnitude: a thin race can be flipped by a small discrepancy, while a comfortable race absorbs a large one. It is judged per contest, not per comparison, since a contest’s outcome depends on the aggregate of its unresolved discrepancies.

To date, every discrepancy AC has found — across deliberately limited coverage — has been non-contest-flipping. A confirmed contest-flipping discrepancy would be the most significant result in the project’s history, and it would take priority over all other work until resolved.

The explanatory states: likely attributable and apparent

When review confirms a real difference, AC forms a hypothesis if it can. If AC holds a specific, benign hypothesis about the vote-reporting logic that would account for the difference, the discrepancy is likely attributable — a hypothesis of convenience, not a verdict. The label says only that AC has a credible explanation it expects to confirm; the resolution still depends entirely on the evidence. Many likely-attributable discrepancies are confirmed by AC’s own checks (a write-in equivalence check, a confirmed ballot-jam re-scan); others need a word from the elections office.

If review produces no benign hypothesis — the difference could be innocent error or it could be fraud, and AC cannot tell — the discrepancy is apparent. Because the cause is genuinely unknown, getting an answer is urgent. Apparent is the only gateway to a confirmed reporting error: AC never reaches a true discrepancy without first being unable to explain the difference.

The number and type of likely-attributable discrepancies are reported, because the pattern is real insight into how a jurisdiction reports votes and helps others know what to look for. Routine initial non-matches are not reported; discrepancies are.

Asking the elections office — a cooperative, weighted step

When a discrepancy needs information AC does not have, AC asks the elections office. This is cooperative, and AC is careful never to call it “escalation” — that word is reserved for the court stage described below. The weight of the ask is proportional to how unexplained the discrepancy is. A likely-attributable discrepancy gets a lighter ask: AC holds a hypothesis and simply wants it confirmed. An apparent discrepancy gets a heavier ask: AC has no explanation, so the inquiry is more formal and more pressing. Either way the question is factual and respectful — “We found a difference of X votes for this contest in this precinct; our transcribed value is Y, the official value is Z; can you explain it?” — and AC gives the office a reasonable time to respond.

A lighter ask resolves to a match if the office confirms, or drops the discrepancy to apparent if the exchange yields no satisfactory resolution. A heavier ask resolves to a match if the office gives a satisfactory explanation, to a true discrepancy if the office acknowledges the error, or — if the exchange yields no satisfactory resolution — leaves the discrepancy ready to escalate.

Match (plausible explanation) — when a credible account is enough

Not every likely-attributable difference is confirmed to the letter, and not every one needs to be. When a difference has a plausible benign explanation — from AC’s own reasoning, a lighter ask, or the official record — that AC accepts without full per-value confirmation, and the difference is non-contest-flipping and of a magnitude the analyst judges benign, it resolves to a match (plausible explanation): the weaker sibling of the arithmetically proven PEWS match. This is the same proportionality that governs spot-checking rather than re-verifying every transcription — AC does not spend its own or an elections office’s finite time chasing certainty on a low-stakes difference that already has a credible account. Two disciplines keep it honest: the explanation must be credible on the evidence and consistent in magnitude with the mechanism it names, and every such match is itemized in the report with its numbers, its mechanism, and the analyst’s recorded rationale, and stays revisable — a reader with the same data can reopen any one. A match (plausible explanation) counts as a match, never a discrepancy, in summary statistics; the Wake County 2022 analysis’s 79 are the canonical worked example (see its investigation walkthrough). A contest-flipping difference is never resolved this way — a plausible story is not enough when an outcome could turn on it, and it gets the heavier ask instead.

Escalatable, escalated, and true discrepancy

An apparent discrepancy that neither AC’s own work nor the elections office could resolve becomes escalatable: AC has no choice but to hand it to a party able to bring a court case — a candidate, a party, a prosecutor, a civil-rights authority. AC escalates only what it cannot explain. (An escalatable discrepancy can still resolve before any handoff — a belated explanation returns it to a match, a belated acknowledgment makes it a true discrepancy — and the court case may simply never materialize, leaving the discrepancy standing, unadjudicated.)

Once a court-capable party takes it up, the discrepancy is escalated — in litigation, with an outcome that depends on the court. If the court finds AV’s evidence correct and the official results are restated, the difference is confirmed as a true discrepancy in the version that contained it. If the court rules the official results correct, AV’s evidence did not persuade. And because AC has never litigated one, some outcomes are genuinely unknown — the model says so rather than pretending otherwise.

A true discrepancy is a discrepancy that the responsible authority — the elections office, by acknowledging it, or a court, by ruling on it — has confirmed to be a genuine vote-reporting error. AC never self-declares a true discrepancy; both routes converge on the same confirmation. AC says the official results were restated, never “corrected”: correction would imply AC fixed every error, when in fact it corrected only the specific errors it found — others may remain in places AC did not check.

Official-results versions, and “true discrepancies that ever existed”

Official results are not a single fixed object; they come in versions — an unofficial election-night tally, later amendments, the certified result, and, rarely, a post-court restatement. AC evaluates its comparisons against each version, and a true discrepancy is always relative to a version. A reporting error confirmed against the election-night numbers, then restated before certification, was a true discrepancy in that earlier version and a match against the certified one.

This is why the honest integrity headline is the number of true discrepancies that ever existed in an analysis, not merely the number standing in the certified results. To date that number is zero: every difference AC has surfaced has been either an artifact of its own process or a real difference that resolved to a match. When it is ever not zero, the report will say so plainly — an error that existed and was put right is a transparency success worth describing in full, not a number to bury.

Lineage and how an analysis reports its discrepancies

A comparison’s current status is tracked separately from its lineage — the story of how it got there. Two comparisons can both be matches while one was an exact match from the start and the other was an apparent discrepancy that an elections office acknowledged and restated. The current status is “match” for both; the lineage is where the second one’s story lives, and a match with a non-trivial lineage is exactly the kind of thing a report should narrate.

At the level of a whole analysis, the everyday integrity headline — when no true discrepancy exists — is built on the likely-attributable discrepancies: how many were found, of what types, and how they resolved. A discrepancy that began as likely-attributable and later became apparent is counted by its current status, with its origin preserved in lineage.

Never “fraud”

AC identifies symptoms, not motives. A discrepancy is a discrepancy regardless of whether it was caused by a clerical error, a software bug, a procedural shortcut, or deliberate manipulation. AC’s methodology detects the fact of the difference; it cannot determine the cause of the difference. Characterizing a discrepancy as “fraud” would be an interpretive leap beyond what the evidence supports.

This discipline — saying only what the tapes show — is the conservative-voice principle stated in full in the Philosophy chapter. In the reporting context it is also essential to AC’s legal positioning: a finding of “apparent discrepancy” is a factual observation that can survive legal scrutiny, while a finding of “fraud” is an accusation that requires proof of intent.

12.4 — Publication, Communication, and the Analysis Window

Where reports are published

AC reports are published on the America Counts website (americacounts.us). Static analysis reports are hosted as permanent pages with stable URLs. Rolling operational logs are maintained as live-updating pages during active analysis and are frozen in place once the analysis concludes. Both formats link to the Actual Vote Video Archive (av.democracycounts.org) for the underlying evidence — readers can follow links from a finding to the specific submissions that support it.

The Actual Vote Video Archive itself serves as a publication channel for the raw evidence. Every approved submission is publicly accessible, with its permanent URL, metadata, and video. The archive is the canonical reference for the evidentiary foundation of any finding.

How AC communicates findings

AC communicates findings to different audiences through different channels, with different levels of detail and different timing considerations.

To elections offices: AC raises discrepancies it cannot resolve internally with the relevant elections office through a cooperative ask — direct correspondence, typically email or phone. The communication is factual, respectful, and specific: “We found a difference of X votes for this contest in this precinct. Our transcribed value from the poll tape is Y; the official reported value is Z. Can you explain the difference?” The tone is collaborative, not accusatory (consistent with the cooperative-ask guidance in the Methodology and Operational Practice chapters), and the weight of the ask is proportional to how unexplained the discrepancy is. AC gives the office a reasonable time to respond before publishing. Contacting the office is an ask, not an escalation; AC reserves “escalation” for handing an unresolved matter to a court-capable party.

To media: When analysis findings are newsworthy — particularly when discrepancies are found or when the analysis covers a politically significant election — AC may communicate findings to journalists. The communication emphasizes the factual findings and the methodology, not political interpretation. AC does not pitch its findings as sensational; it presents them as evidence that journalists can contextualize for their audiences.

To advocacy organizations and partners: AC shares findings with partner organizations (Scrutineers, local civic groups, voting rights organizations) that can use the evidence in their own work. The sharing is governed by AC’s non-partisan commitments — findings are shared with organizations on all sides of the political spectrum, and AC does not allow any partner to claim exclusive access to findings.

To the general public: Through the published reports and the Actual Vote Video Archive, findings are available to anyone with internet access. No paywall, no registration requirement, no special credential. The public accessibility of findings is a core expression of the Transparency Principle (Section 13.1).

Timing considerations

The timing of publication depends on the Analysis Window. Before-certification findings have operational urgency: they may inform corrections, recounts, or challenges, and they must be published quickly enough to be actionable. AC’s working principle is to publish before-certification findings as soon as they are ready, even if the analysis is incomplete, rather than waiting for completeness and missing the window.

After-certification findings do not have the same urgency but still serve important functions. They contribute to the permanent record, inform future analyses, support transparency, and may provide evidence for legal challenges that proceed after certification. After-certification publication can afford to be more thorough and more carefully crafted.

The Analysis Window and publication timing

The Analysis Window is one of the most important framings in AC’s reporting because it shapes what an analysis can accomplish and how its findings should be understood.

Every US election has a certification date — a fixed date by which election officials must formally declare the results final. This date divides the analytical world into two windows. The before-certification window runs from election day until the certification date. Findings produced in this window have operational urgency: they can in principle inform corrections, recounts, or challenges before results become final. The after-certification window runs from the certification date onward. Findings in this window cannot affect the certified results directly — changing them requires formal legal proceedings — but they serve transparency, the historical record, and future elections.

The window shapes AC’s decisions about which analyses to undertake, how fast to move, and how to present findings. Before-certification analyses must accept that they cannot be exhaustive — the window is too short for AC’s resources, so AC focuses on the most important parts (the most contested races, the largest jurisdictions, the most accessible recordings) and accepts that the rest is out of scope. AC’s working principle is to publish before-certification findings as soon as they are ready, even if the analysis is incomplete, rather than waiting for completeness and missing the window.

After-certification analyses can be more thorough. Without deadline pressure, AC can take more time per investigation, follow up on apparent discrepancies more carefully, and produce more rigorous findings. After-certification findings contribute to the permanent record, inform future analyses, support transparency, and may provide evidence for legal challenges that proceed after certification.

Many analyses straddle the two windows. AC’s typical approach is to publish whatever findings are ready by the certification date, continue the in-progress work afterward, update the report when new findings arrive, and be explicit about which findings were available before certification and which were added after. The 2024 General Election News Page is an example of this kind of straddle: the page was active during the before-certification window in November–December 2024 and continued to receive updates in the after-certification window.

Every Actual Vote analysis report should include the Analysis Window in its front matter. For analyses that straddle both windows, the front matter should note both phases.

The rolling-update model

Reports are not always static documents published once and never touched again. The rolling-update model — exemplified by the 2024 General Election News Page — allows findings to evolve as new information arrives. An entry posted during election week may be updated a month later when an elections office responds to a heavier ask. A finding characterized as an apparent discrepancy may be reclassified as a match when the explanation arrives.

The rolling-update model trades the definitiveness of a static report for the accuracy of a living document. For complex, multi-state analyses that unfold over weeks or months, the rolling model is often more appropriate. For bounded, completed analyses, the static model provides the clarity and citeability that downstream users need.

Versioning and updates

Static reports may be updated after publication when new information arrives. Common reasons for updates include a late-arriving response from an elections office that explains an apparent discrepancy, a correction to a transcription error discovered after publication, a new submission that arrived after the report was published but pertains to the analysis, or a clarification in response to reader questions. Updates should be disclosed: the report should carry a version number or date stamp, a changelog section listing what changed in each version, and the original published version should be preserved (not overwritten) so that anyone who cited the original can still find it. The AC Findings Changelog (B.8) is the canonical place for tracking changes across all published analyses.

What happens after publication

In the days and weeks after a report is published, AC distributes the report through its website and communication channels, notifies relevant elections offices (especially where unresolved discrepancies were raised), notifies the volunteers whose recordings supported the analysis, and responds to any media inquiries factually, pointing reporters to the report itself rather than offering commentary that extends the report’s claims.

Over the medium term, late-arriving information may update the analysis (new submissions, late responses from offices, new context). Cross-analysis learning feeds into future work: per-jurisdiction knowledge grows, the discrepancy code list expands, the state-by-state appendix gains entries. Relationships with cooperative elections offices are maintained for future cycles. Volunteers who participated remain in AC’s network.

Over the long term, each analysis becomes a data point in an accumulating body of evidence. As AC publishes more analyses with consistent methodology, the track record itself becomes evidence about the state of vote reporting in the United States. The repeated finding that careful comparisons turn up no surviving discrepancies is itself a substantive claim about how elections are run. A growing knowledge base about specific jurisdictions makes future analyses faster and more accurate. A growing demonstration of the methodology provides proof of concept that independent vote reporting verification is feasible.

12.5 — Interpreting a “Clean” Result

What it means when zero discrepancies are found

The most common headline finding in AC’s published analyses to date has been a clean result: zero true discrepancies after thorough investigation. The NC 2024 Primary Report, which covered 321,502 individual votes across seven counties, is the canonical example.

A clean result means that for every comparison AC performed — every County × Precinct × Voting Method × Contest × Choice combination for which AC had both a transcribed poll tape value and an official reported value — the values either matched exactly or differed for an explained reason. The reporting layer was accurate for the portions of the election that AC analyzed.

This is a real, substantive finding. It tells the reader something meaningful about the accuracy of vote reporting in the analyzed jurisdictions. It provides independent confirmation — from a non-partisan organization using primary-source evidence — that the elections offices got the reporting right, at least for the specific precincts and contests covered.

What it does not mean

A clean result does not mean that the entire election was clean. It means that the specific portions AC analyzed were clean. This is the coverage caveat: AC’s analysis covers the precincts for which AC has recordings, not the full jurisdiction. A clean finding for 52 precincts in seven counties is not a finding about the other precincts in those counties, or about other counties in the state, or about other states.

A clean result does not mean that there are no problems with the election as a whole. The matched poll tape caveat (Section 13.4) applies: a clean result confirms that the reporting layer was accurate, but says nothing about whether the counting layer was accurate. Problems upstream of the poll tape — voting machine errors, ballot marking problems, voter registration issues — are outside the scope of what AV measures.

A clean result does not mean that future elections will also be clean. Each election is an independent event. Clean findings in 2024 do not guarantee clean findings in 2028. The value of AV is precisely that it checks each election independently rather than assuming accuracy based on past performance.

Why clean results are valuable

Clean results are sometimes perceived as boring or uninformative — “you found nothing, so what’s the point?” This perception misunderstands the nature of verification.

Clean results build confidence. In an environment where public trust in elections is under strain, independent confirmation that vote reporting was accurate is valuable information. A reader who was uncertain about the accuracy of their local elections can see, in a published report backed by primary-source evidence, that an independent non-partisan organization checked the numbers and found them correct. That confirmation is a public good.

Clean results validate the methodology. Each clean analysis demonstrates that the AV pipeline works: that ordinary citizens can collect evidence, that the evidence can be transcribed and compared, that the comparison produces meaningful results, and that the results can be published transparently. The methodology’s credibility grows with each successful application.

Clean results create a baseline. As AC accumulates clean analyses across elections and jurisdictions, a baseline picture of vote reporting accuracy emerges. If a future analysis in the same jurisdiction finds discrepancies, the baseline provides context: is this a new problem, or a change from a previously clean pattern? Baselines are valuable for detecting trends and anomalies.

Clean results contribute to the deterrent effect. Elections offices and other actors in the vote reporting process know that independent observers are checking the numbers. Even when the checking reveals no problems, the knowledge that checking is happening creates an incentive to maintain accuracy. The deterrent effect does not require finding problems — it requires the credible possibility of finding them.

Clean results are the expected outcome. Most elections in the United States are administered conscientiously by honest professionals. The expected result of checking the reporting layer is that it was done correctly. AV’s repeated clean findings are consistent with this expectation. If AV frequently found discrepancies, that would be alarming. That AV frequently finds accuracy is reassuring — and the reassurance is evidence-based, not faith-based.

12.6 — Reasoning Under Incomplete Documentation

Every number in an AV report rests on a model of how a particular jurisdiction reports votes. Part of that model is documented in statute, rule, and poll worker manuals; part of it is inferred. This section states the epistemic position behind AV’s findings plainly, because the credibility of the work depends on being precise about what AC knows versus what AC has reasonably concluded.

The position: a constraint of access, not a verdict about counties

For all AC knows, a county’s internal files may contain complete documentation of its reporting process. From the outside, it is not possible to tell. Statutes give the legal skeleton, administrative rules add operational detail, and poll worker manuals add procedure — but the end-to-end function that maps what a tabulator prints at the close of polls to what appears in the official export weeks later, including every canvass adjustment, write-in resolution, merged precinct, re-scan, and redaction rule, is only partially visible from the public record. AC’s work across multiple states indicates that some of this process is formally documented and some is not; how much internal documentation exists, and how much of the process it covers, is genuinely unknown to an outside observer. The only way to find out would be to ask every elections office for everything — an inappropriate request, and contrary to AC’s standing policy of not badgering election administrators (Section 8 and Section 12.4). AC presumes elections offices are proceeding appropriately under status quo conditions, with internal procedures, vendor documentation, and training doing work the public record does not show. The defining constraint is access, not a verdict about what counties have written down.

This produces two failure modes to navigate between. An auditor who demands the complete specification before comparing anything will either never compare anything or become a nuisance to the very offices whose cooperation the work depends on. An auditor who assumes their own model is correct, without distinguishing what they verified from what they assumed, will eventually publish a false alarm — and in election work, a false alarm is not a neutral error; it feeds narratives that outlive every correction. AC’s methodology lives between these, and the discipline that keeps it there is abductive reasoning, practiced openly: inference to the best explanation, with the inference stated rather than hidden, and revised when new information warrants.

The standard: over-determination, not mere plausibility

An assumption earns its place in an analysis when it is the only reading that makes common sense, is consistent with all the data held, and survives a deliberate attempt to break it. The property that separates rigor from rationalization is over-determination: a sound inference is not merely consistent with the observations, it is forced by them, while the alternatives are not. When an aggregation hypothesis must reproduce the exact arithmetic structure of dozens of independent rows simultaneously, the set of wrong hypotheses that happen to reproduce it shrinks toward zero. The strength of an AV inference is the number of independent constraints its explanation satisfies at once, and the implausibility of any rival explanation satisfying them all by coincidence. Three cases illustrate.

One-stop aggregation (Wake County, 2022). North Carolina reports early voting under site-and-range identifiers; no public document AC located specifies what they aggregate. Per-machine tapes ran systematically short of the official totals across every one-stop site. The hypothesis that the pseudo-precincts aggregate all of a site’s machines had to reproduce the exact structure of the shortfalls at once — two-machine sites short by machine-sized amounts, single-machine sites matching — and a wrong hypothesis would have left residue somewhere. The abduction came first; it then identified the single question worth asking the county, whose DS200 certification list converted the hypothesis into documentation (Section 18 and the Wake County 2022 analysis).

Operational events (Bat Cave, Henderson County, 2024). A polling place’s results were irreconcilable against the law and the manuals. This was the case where the data did not over-determine the answer — several stories could have fit the gap, and the true one was a contingent operational event (a weather-driven on-the-fly polling place merger) that no amount of analysis would have produced. That is precisely when AC asks the county rather than guesses, and precisely why an unexplained mismatch is an apparent discrepancy and not an accusation. The space of legitimate operational events is larger than the space of publicly documented ones, which cuts both ways.

Write-in representation (the plausibly equivalent write-in situation, 2024). Where a single contest’s difference reconciles exactly as raw-versus-resolved write-in accounting, the over-determination is layered: the within-contest arithmetic closes, the tape’s own internal checksum passes, and (where dated official captures exist) the election-night data matches the tape before the canvass adjustment. For a benign coincidence to fake all of these at once is not credible. When such an abduction recurs across analyses, AC promotes it from case-by-case judgment to a named, documented rule — as it did with the write-in equivalence check (Section 8.7).

Calibrated humility

Two opposite mistakes are available here, and the discipline is to make neither.

The first mistake is overstating doubt. Most of what an AV analysis does is simple enough that AC can be practically certain it is right — the arithmetic is plain, the structure is obvious, and a competent reader would reach the same reading. Hedging such cases to death would be its own form of dishonesty and would bury genuinely strong results under false modesty. When a situation is simple, AC treats it as settled.

The second mistake is the one experience teaches you to fear. In a complicated situation — not even a maximally complicated one, just an ordinary messy one — something that passes a normal common-sense check can quietly hide deep, subtle, undocumented logic. Everything looks fine for now. Then, downstream, a strange result surfaces, and tracing it back lands on an assumption that seemed entirely reasonable at the time, by any normal standard, while a dozen other assumptions of the very same strength, made in the very same analysis, caused no trouble at all. Afterward there is always the observation that “you should have directly checked the thing that turned out to be responsible.” But one cannot directly check a thousand things. AC uses experience to narrow a thousand candidates to the handful most worth checking, and checks those — and that triage, however good, is never perfect. It would be far better if all the logic were simply documented, but that standard is rarely met outside domains where the stakes force it (aviation, or any system whose managers commit the resources because they know it is critical) — which is exactly what AC’s recommendations to elections offices (Section 15.1) push toward.

The honest posture is therefore a paired one: AC is genuinely confident about the assumptions in the AV pipeline — most are simple, over-determined, and battle-tested across analyses — and AC holds real humility about the residual, because the real world is messy and humans are imperfect, and it can never be fully ruled out that a real error or attack, accidental or deliberate, left the numbers looking exactly like something that makes sense. A match under an assumed aggregation rule is evidence the rule is right and the reporting is right; it is not proof that nothing went wrong in a way that happens to mimic correctness.

The risk is not symmetric, and knowing which way it leans is most of the wisdom. The dangerous failure is the apparent match that hides a real discrepancy — quiet, with no alarm bell and nothing in the output to flag it — and that risk is magnified at small sample sizes, because when an analysis rests on two or three precincts in a county, the chance of everything appearing to match (whether or not it truly does) is simply higher. Apparent discrepancies are the easier side: AC’s experience with elections offices is that the ones with a good underlying explanation are usually resolved without much difficulty, through a single cooperative exchange. So AC’s error budget is lopsided toward the quiet false match, and that is where its humility, its over-determination standard, and its insistence on re-runnable analyses are all concentrated. This is why the over-determination bar is not optional decoration — an assumption that merely lets the numbers tie is insufficient; the bar is an assumption forced by enough independent observations that a wrong version would have left a visible trace. And it is why the revision mechanism is load-bearing: the cases where a tidy match was secretly wrong are exactly the ones a later piece of information is positioned to catch — if the analysis was built to be re-run when that information arrives.

The standing assumptions AC carries

The principle becomes concrete in the specific assumptions an analysis makes when the public record runs out. These are cross-jurisdictional methodology assumptions, not a single state’s quirks; each is adopted because confirming it through primary legal research would be a large undertaking AC prioritizes rather than performs exhaustively, and each is stated with its basis and its failure mode, because the failure mode is the honest part. Going forward, each state’s research pass produces its own short version of this list as part of its knowledge notes; the basis of any entry can be upgraded when primary-source research or a county conversation turns an inference into a documented fact.

Multi-tabulator precinct aggregation. Where a precinct runs more than one tabulator — uncommon in North Carolina, where one machine per election-day precinct is typical, but routine in states like Florida — each machine prints its own results tape and the precinct’s official total is the sum of those tapes. Basis: common-sense inference, confirmable per precinct on demand. No document states “add the tapes,” and none tells AC in advance how many tabulators a given precinct ran. The failure mode is concrete: a precinct’s official total can look like four tapes’ worth of votes while AC recorded three. The resolution is the model for the whole approach — AC asks the county the one narrow question (how many tabulators were at this precinct?), not for the full equipment layout, which would be an inappropriate ask. If the answer is four, the county can sometimes furnish the missing tape later in the chain of custody; if it cannot, there is simply no comparison there; and if it does not answer, AC cannot rule out that three tabulators were in play and a genuine error or attack left the totals about one tape short.

The comparison grain itself is correct. That summing transcribed values to a particular grain and matching them against an official value is the right aggregation for this jurisdiction and election. Basis: data-science inference, strengthened or weakened by sample size. The 2024 Georgia General analysis showed this is not trivial: nothing told AC how Georgia’s early-voting results redistribute to home precincts or how its pseudo-precincts compose. AC found an aggregation under which the numbers looked like correct reporting and inferred the grain was right — and cannot rule out that a reporting error merely made it look correct while the true rule is something else. With many precincts the risk shrinks toward nothing; with two or three precincts in a county it remains meaningfully abductive.

One-stop aggregation in North Carolina. That NC’s site-and-range one-stop pseudo-precincts aggregate all of a site’s machines, so a comparison requires every machine’s tape (the Wake decode in “The standard,” above). Basis: backed out from data; county-confirmed for one case; assumed and revisable elsewhere. On the information available it is hard to see how else it could work — but two live failure paths remain: NC could change its one-stop aggregation in a future election, so logic AC assumes fixed silently goes stale and real discrepancies are missed; or the one-stop tapes AC does not have could run on a different system it has no information about, in a way that hides an error. Given how much election machinery and procedure varies, neither is far-fetched.

An absent official value means zero or not-applicable — not a withheld or failed report. Basis: over-determined by NC’s format consistency, but known to be falsifiable. This is flagged falsifiable because AC has watched it fail: in Georgia 2024, roughly seventy percent of recorded values had no published precinct-level counterpart — neither zero nor not-applicable, but unpublished — and a blank can also be a small-count anonymity redaction. AC treats absent NC election-day values as zero or not-applicable, holds that as an assumption rather than a fact, and tests it when the pattern of blanks looks structured rather than incidental.

Terminology on the tape and in the official results refers to the same things, after reasoning. That differently-worded lines reconcile — a tape’s “Absentee” and “Provisional” sections mapping onto official method blocks, or a candidate spelled one way on the tape and another in the export. Basis: human-reviewed inference, partly mechanized as fuzzy matching. The county has internal logic for these translations; outside observers reconstruct it, and AC’s pipeline does so with a reviewed matching step. The genuine hazard is an ambiguous case where one plausible matching yields perfect agreement and another yields apparent discrepancies, leaving AC to judge how much to trust the clean reading. This is the failure AC’s recommendations to elections offices are meant to prevent (Section 15.1, items on consistent naming and documented reporting logic).

Write-in resolution (the plausibly equivalent write-in situation). That a tape’s raw write-in count and the official results’ canvass-resolved write-in lines describe the same votes (Section 8.7). Basis: informal confirmation from election-staff conversations, plus over-determined arithmetic; the ask scales with magnitude. The pattern is inherently plausible — voters do write in candidates already on the ballot, forcing exactly this situation — and where the arithmetic closes exactly, the tape checksums internally, and dated captures show the election-night figure matching the tape before the canvass adjustment, AC treats it as a match. Where the magnitude is large enough to matter, it warrants a call rather than an assumption.

The official snapshot compared against is the one AC intends. That the official-results file AC captured corresponds to the canvass or certification state it means to audit against, not a superseded version. Basis: provenance and dated capture. Official results evolve between election night, canvass, and certification — Georgia 2024 made this vivid, where a precinct’s presidential value changed by one vote between the November 9 and November 18 exports, and that evolution is what corroborated the write-in resolution. “Which version of the official results” is therefore itself an assumption; AC mitigates by capturing results with timestamps and provenance and stating in each report which official state the comparison used.

A precinct present in official results but absent from AC’s recordings is a coverage gap — not a merger, suppression, or relocation — until investigation shows otherwise. Basis: default assumption, with anomaly as the trigger to ask. This is the assumption the Bat Cave case (in “The standard,” above) exists to puncture: a precinct’s votes lawfully appeared elsewhere because a flooded road forced an on-the-fly polling-place merger, completely legitimate and underivable from any document. The default (“AC simply does not have that tape”) is held only until the data looks strange enough to warrant the email.

A candid caveat: the no-badgering policy

One governance point deserves stating plainly, because a critic would otherwise state it first. AC’s policy of not pestering elections offices does double duty: read charitably it is courtesy and stewardship of a finite relationship, but read uncharitably it could become a way to avoid learning inconvenient facts. AC answers this with a bright line the record bears out — the policy gates trivia, never findings. Any pattern that survives internal investigation, and anything that could affect an outcome, is raised with the office, politely and on the record. The Wake County 2022 analysis is the proof: the patterns that mattered produced an ask and a same-week reply, while one-and-two-vote write-in items that could not have changed any outcome were characterized in the report as likely attributable and left unsent. The asymmetry is the safeguard.

What this means for reports

The honest formulation of any AV result is conditional: under AC’s model of the jurisdiction’s reporting process — most of it documented, some of it inferred and stated — the comparison produced so many exact matches, likely attributable discrepancies, and apparent discrepancies. AC is not mathematically certain that no discrepancy hides inside an inference; a match under an assumed aggregation rule is evidence the rule is right and evidence the reporting is right, and AC says so. Three practices keep this honest rather than hand-wavy. Assumptions are enumerated, not ambient — each analysis records what was inferred (aggregation grain, identifier reconciliations, treatment of unpublished values) in its manifest and methodology notes, so a reader who wants to attack the conclusions knows where to aim. The vocabulary carries the epistemics — the discrepancy vocabulary of Section 12.3 exists so a reader can tell, for every number, which kind of knowing stands behind it. And everything is revisable, cheaply — frozen inputs, versioned reports, and re-runnable pipelines mean that when the next operational explanation arrives, updating the analysis is routine. Abduction without a revision mechanism is guessing with confidence; abduction with one is how empirical science works.

None of this is a compromise of rigor. Certainty was never available to an outside auditor under status quo conditions, and demanding it would make AC both a worse neighbor and a worse auditor. In an ideal world, every jurisdiction would publish a specification complete enough to make these inferences unnecessary, and AV reports would be shorter for it. Until then, stated assumptions, disciplined and over-determined abduction, and open revision are what rigor looks like from where an honest outside observer actually stands — and they are what allows a clean AV result to mean something, because a confirmation from an auditor who assumed everything would be a confirmation from no one.