G: Yes. The biggest problem is exactly the one you anticipated: the United States does not maintain a national count of fraudulent ballots**, and “fraud cases” are not interchangeable with “fraudulent votes.” So I think the cleanest comparison is to estimate everything in the same unit: presidential votes/ballots from 2004 through 2024, thereby excluding Bush–Gore 2000 as you requested.

For actual voter fraud, I would use about 10,000 fraudulent ballots as a reasonable working number over that period, with something like 2,500–20,000 as a defensible uncertainty band. Here's where that comes from. Studies of elections that were intensively investigated have produced fraud rates around 0.0003%–0.0025%; Brennan cites, for example, 0.0009% in the extraordinarily scrutinized 2004 Washington election and 0.00004% in Ohio. citeturn434983search2turn434983search4 There were roughly 840 million presidential ballots cast from 2004 through 2024. Applying a middle working rate of about 0.001% gives ~8,400 fraudulent votes, which I'm comfortable rounding to ~10,000 for what you're doing. This is explicitly an extrapolation, not a measured national count.

As a reality check from the opposite direction, Heritage—the organization whose database is specifically intended to document election fraud—currently has 1,467 proven election-fraud cases dated 2000 or later, out of 1,619 total entries. Heritage itself says the database is not exhaustive. Crucially, those are cases, some involving one ballot and some involving multiple ballots, so they don't give us the numerator you need. citeturn302847view0turn434983search15 And in-person impersonation, which gets a disproportionate amount of attention, is microscopic: Levitt found only 31 credible incidents from 2000–2014 among more than one billion ballots, some of those allegations not even ultimately proven. citeturn302847search0

Now the really interesting comparison.

Election researchers measure residual votes: ballots on which no valid presidential vote was recorded because of an undervote, overvote, or similar failure. It isn't synonymous with voter error, because some people intentionally leave the presidential race blank. But surveys put intentional presidential abstention at roughly 0.3–0.7% of voters. citeturn400895search8 The observed residual rate has generally hovered around 1%: about 1.06% in 2004, approximately 1% in 2008, 0.99% in 2012, and 1.39% in 2016; it fell again in 2020. citeturn400895search7turn248007search48turn248007search8turn481647search4

That gives us a surprisingly useful back-of-the-envelope calculation. Across ~840 million presidential ballots:

roughly 8–9 million presidential votes were residual votes;
perhaps 2.5–5.9 million of those were intentional abstentions;
leaving roughly 2.5–6 million unintentionally lost/miscast presidential votes;
I'd use ~4 million as the working midpoint.

And there is strong evidence that a substantial portion really is caused by the interface/ballot system rather than the voter merely changing his mind. In 2004, changing voting equipment dropped residual-vote rates dramatically; the Caltech/MIT work estimated that reforms between 2000 and 2004 alone recovered roughly one million votes. citeturn400895search3turn400895search7 In one especially clean example, North Carolina had a confusing straight-party rule: after voter education designed specifically around that problem, presidential residual votes fell from 2.2% in 2004 to 1.0% in 2008. citeturn248007search4 And in Florida, simply changing an instruction telling voters that contests continued onto the back of the ballot was associated with 28,000 fewer lost votes in the affected ballot questions. citeturn248007search1

So if you force me to divide that ~4 million into your two buckets, I'd give you this low-confidence but numerically reasonable decomposition:

2004–2024, presidential contest only	Working estimate
Fraudulent votes	~10,000
Wrong/lost because of ballot, interface, or voting-system design	~2–3 million
Personal voter mistakes / “brain farts” independent of obvious design problems	~1–2 million
Design + ordinary mistake combined	~4 million

I have substantially more confidence in “~4 million unintended errors” than in the 2–3m / 1–2m division. There simply isn't a national dataset that observes a person and tells us why they produced a residual vote. The partition is an inference from natural experiments comparing different machines, layouts, instructions, and jurisdictions.

But for the comparison I think you're trying to construct, you actually don't need that partition. The defensible statement is approximately:

Since 2001, ordinary voting mistakes and voting-system/ballot-design problems have probably caused a few million presidential votes not to reflect the voter's intention, while actual fraudulent voting is probably measured in the thousands or low tens of thousands.

Using the midpoint numbers, that's about 4,000,000 versus 10,000 — roughly 400 to 1.

And that's deliberately conservative in one important way: I counted only the presidential line of the ballot. Down-ballot contests and ballot propositions have dramatically higher residual-vote rates—one 2004 study found a median 6.5% residual rate on ballot issues versus 0.9% for president. citeturn400895search6 I did not add those, because then we'd be counting multiple choices per ballot on the error side while ordinarily counting one fraudulent ballot on the fraud side, which would make the comparison mathematically uglier.

So ~10,000 vs ~4 million is the pair of numbers I'd use, accompanied by “both are estimates; the error number is far better measured than the fraud number.”

G: Yes—that suspicion is exactly the right instinct here. The useful version is not “I found a ratio that supports the point,” but “I can tell you what each numerator actually contains, how it was assembled, and where the estimate starts becoming inference.”

I’m going to tighten this around three things: Heritage’s case-by-case sourcing and what fraction can be converted into actual ballots/votes; the best empirical literature on residual votes and ballot-design effects; and a range that survives someone pushing hard on definitions rather than relying on one convenient midpoint.

G: [Non-text content: thoughts; content_type=thoughts]

G: The Heritage digging changes one important part of my previous answer: I would withdraw the “~10,000 fraudulent votes” midpoint as insufficiently defensible. It was a plausible order-of-magnitude extrapolation, but it is not something Heritage’s data actually allow us to calculate.

And there’s a very concrete reason. Heritage’s records mix together single illegal ballots, multi-ballot schemes, attempted fraud, false registrations, petition-signature fraud, and multiple defendants from the same ballot scheme. So even manually summing every number mentioned in its narratives would sometimes count the same affected ballots several times. That is exactly the sort of hidden flaw someone could legitimately use to knock down the comparison.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes. After digging into it, I’d substantially tighten the way I would present the comparison—and I would not use my earlier “10,000 fraudulent votes” as though it were an established number. That number was too inferential for the role you want it to play.

The strongest starting point is Heritage itself. Its live database currently shows 1,620 entries. The archived December 12, 2025 dataset contains 1,619 entries, of which 1,467 are dated 2000 or later. Of the 1,619, 1,381 are criminal convictions; the rest are diversion agreements, civil penalties, judicial findings, or official findings. Heritage explicitly describes the database as a sampling, not a comprehensive census. citeturn915806search3turn915806search1turn184130view1

And Heritage does appear to be doing genuine case research rather than just throwing newspaper allegations into a list. It says it uses court records, reported cases, Westlaw, government records, and press accounts; individual entries ordinarily contain citations or case numbers. One Virginia entry, for example, identifies the actual criminal case number. Others provide multiple underlying sources. citeturn915806search12turn910429search0

So I would regard “about 1,500 proven election-fraud cases since roughly 2000” as a perfectly legitimate datum.

But here's the extremely important catch.

Heritage's 1,467 cases cannot be converted directly into 1,467 fraudulent votes

The entries aren't homogeneous units.

One Heritage entry is essentially one fraudulent ballot. Richard Dohmen tried to vote both his ballot and his deceased wife's ballot. citeturn910429search0

Another case involved three ballots. Cynthia Gonzalez pleaded guilty to three election-code violations after marking and mailing ballots belonging to other people. citeturn910429search7

Another involved 24 counts of voter fraud. citeturn910429search5

Then you get cases such as the 2018 Gregg County, Texas operation. Heritage has separate database entries for Shannon Brown, Marlena Jackson, Charlie Burns Jr., and DeWayne Ward. But all four records describe the same scheme involving more than 100 absentee ballots. If we mechanically read each Heritage entry and assigned “100 ballots” to it, we'd count those same ballots roughly four times. citeturn910429search2turn910429search3turn910429search6turn910429search10

And some entries aren't fraudulent votes at all. Heritage includes things such as false voter registration and fraudulent candidate-petition signatures. Diana Watt, for instance, forged signatures on candidate qualifying petitions; that belongs in an election-fraud database, but there is no fraudulent ballot to add to our “wrong votes” numerator. citeturn720629search3

There's the opposite problem too: some schemes affected an unknown number of ballots. In North Carolina's 2018 Ninth Congressional District election, seven Heritage entries stem from the same absentee-ballot operation, but investigators concluded that the operation's tampering and concealment made the exact number of affected votes impossible to reconstruct. citeturn184130view1

So the Heritage database gives us something much more defensible than I initially realized, but also something narrower:

1,467 documented instances/defendants/findings since 2000, not 1,467 votes and not some calculable X thousand votes.

There is no honest multiplication factor.

There's a second independent dataset that makes me considerably more comfortable with that conclusion. In 2012, the Carnegie-Knight News21 project attempted something Heritage explicitly does not claim to do: reporters sent more than 2,000 public-records requests to elections officials and prosecutors in all 50 states and reviewed nearly 5,000 court documents, official records, and media reports covering 2000–2012. They found 2,068 alleged election-fraud cases of all sorts. citeturn623088search1turn623088search5

That number included allegations that ultimately went nowhere, so it isn't directly comparable to Heritage's proven cases. But that's useful: two entirely different collection methods, one explicitly conservative and one exhaustive-by-public-records methodology, both wind up with quantities in the low thousands rather than tens or hundreds of thousands.

News21 found only 10 allegations of in-person voter impersonation during that period. Absentee-ballot allegations were much more common—491—but again, “491 cases” isn't “491 ballots.” citeturn623088search0turn623088search2

And this inability to give you a national fraud-ballot count isn't political hedging. The Government Accountability Office reached essentially the methodological conclusion we're running into: there is no single federal/state/local database covering the universe of election fraud, agencies collect different information, and therefore the complete incidence cannot be determined from existing records. citeturn623088search3

So if somebody challenges you with “How many fraudulent votes were there?”, the technically strongest answer is:

Nobody has a defensible national total of fraudulent ballots. We have case databases, and cases are not ballots.

That is stronger, not weaker, than inventing an estimate.

The voter-error side is rather different because we actually have a measurable phenomenon.

Election researchers use the residual vote rate: ballots cast where no valid vote for a particular office is recorded. That includes both undervotes and overvotes. There are three principal explanations: intentional abstention, voter confusion/error, and technology/interface failure. citeturn944228search11

The Caltech/MIT Voting Technology Project has been studying this since 2000. Its original national work found something particularly useful for your argument: residual-vote rates varied systematically with voting technology. That's the empirical reason we know that not all residual votes are deliberate abstentions. Their historical work estimated that roughly 30% of presidential residual votes were intentional, leaving the rest attributable to error, confusion, or failure to record the intended vote. They explicitly identified ballot/user-interface design as an important source of lost votes. citeturn758620search68

And there's a beautiful natural experiment immediately after 2000.

The national presidential residual-vote rate fell from 1.90% in 2000 to 1.06% in 2004. Charles Stewart estimated that the improvement corresponded to roughly one million recovered votes, associated with new equipment and election-administration changes. citeturn758620search0

That is a much firmer piece of evidence than my previous broad extrapolation. We aren't guessing that design matters: changing the systems measurably caused residual votes to disappear.

The rates didn't simply vanish afterward, either. The nationwide presidential residual rate was 0.99% in 2012 and 1.41% in 2016. Importantly, researchers concluded that most of the increase in 2016 was deliberate candidate abstention rather than worse voting technology, which is exactly the distinction we'd want them to make rather than lazily labeling every residual vote an error. citeturn944228search0turn944228search1

And there are very clean ballot-design examples where you can practically watch the mistake being generated.

In Georgia in 2020, researchers found that a confusing two-column layout in Gwinnett County led roughly 4,200 voters to overvote one U.S. Senate contest, invalidating those votes. That's one county, one contest, one election, with the mechanism identified statistically. citeturn944228search4

In Florida, usability testing found people failing to notice that contests continued on the reverse side of a ballot. Two counties changed the instruction and placement before the election; their residual rate on the first reverse-side contest was 10.2% versus 14.7% in comparable counties that didn't make the change, corresponding to an estimated 28,000 votes saved by changing one instruction. citeturn944228search15

That is extremely useful because it isn't based upon asking somebody afterward, “Did you screw up?” The ballot design was changed, and the error rate changed with it.

So here is where I would now put the numbers if we wanted something that I would personally be comfortable defending under questioning.

Quantity	What we can safely say
Heritage proven election-fraud entries, 2000–2024	1,467
Total Heritage entries through Dec. 2025	1,619
Of those, criminal convictions	1,381
News21 alleged fraud cases, 2000–2012	2,068
Exact fraudulent ballots nationally, 2001–2025	Unknown / not reconstructable from available data
Presidential residual vote rate, 2004	1.06%
Presidential residual vote rate, 2012	0.99%
Presidential residual vote rate, 2016	1.41%
Votes estimated recovered from election-system improvements, 2000→2004 alone	~1,000,000
Documented ballot-design-caused errors in one Georgia county/one Senate contest in 2020	~4,200
Votes saved by one Florida ballot-instruction/design improvement	~28,000

The especially interesting comparison, therefore, isn't actually the 400:1 number I gave you before.

The much harder-to-attack statement is:

Heritage has documented roughly 1,500 proven election-fraud cases over the last quarter-century. We cannot legitimately convert those cases into a national number of fraudulent ballots. But independently, election researchers have measured ballot-system improvements that recovered roughly one million presidential votes in just the four years from 2000 to 2004, and individual ballot-design changes have been shown to prevent thousands or tens of thousands of errors in single jurisdictions.

That is substantially better rhetorically because there is almost no modeling in it.

And if someone responds, “But one Heritage case can contain multiple fraudulent ballots,” the answer is simply yes. That's true. You acknowledge it. But then the reverse problem also applies—several Heritage cases can refer to the same ballots, some entries involve registration or petitions and zero ballots, and some schemes' affected-ballot totals are unknowable. That is why responsible analysis doesn't pretend the database contains a national ballot total. Heritage itself doesn't claim that it does. citeturn915806search1turn910429search2turn910429search3

If you still want a numerical estimate of unintended votes over 2004–2024, I think we can produce one, but I would now treat it as a separate modeling exercise and build a conservative / central / aggressive range from actual election-by-election residual rates rather than the shortcut I used before. That will give you something where every assumption can be laid on the table and challenged individually—which sounds much closer to what you're after.

G: Yes. And phrased that way, I think your argument is both narrower and considerably more defensible than the comparison I was constructing.

You don't actually need to establish how many fraudulent votes occurred over 25 years. You're testing a proposition about relative sources of erroneous election outcomes at the ballot level:

After removing intentional abstentions from the residual-vote rate, are ordinary voter/interface errors sufficiently common that they constitute a larger source of votes failing to reflect voter intent than intentional voter fraud?

The evidence strongly points to yes, and we can establish that without pretending we know an exact national fraud rate.

The key empirical observation is that residual votes respond to the voting system. If residuals were overwhelmingly people deliberately skipping a race, changing ballot technology, layout, instructions, or voter education shouldn't substantially change their frequency. But it does. The 2000→2004 work is particularly strong: the presidential residual rate dropped from roughly 1.9% to 1.1%, with researchers estimating about one million additional presidential votes successfully recorded following equipment and administrative improvements. That's evidence of error reduction, not merely a count of blank ballots.

Now deliberately be extremely conservative. Suppose we say that after removing intentional abstention, only 0.25% of presidential ballots represent mistakes attributable to voters interacting with the ballot/system. That's substantially below the observed ~1% residual rates and deliberately gives away a lot of uncertain territory.

In a modern presidential election with ~150–160 million ballots:

150,000,000 × 0.0025 = 375,000 unintended failures to record a presidential choice.

At 0.1%, an almost absurdly conservative assumption:

150,000,000 × 0.001 = 150,000.

So you don't need my previous estimate of four million over 25 years. The argument can rest on the much more modest claim that a single national election plausibly produces hundreds of thousands of voter-intent errors, even after excluding intentional abstention.

Then compare that with what fraud research finds. Heritage has accumulated roughly 1,500 proven election-fraud entries over approximately a quarter-century—and, as we established, those cannot properly be converted into either 1,500 ballots or some larger national ballot total. Other intensive investigations likewise find fraud to be extremely uncommon relative to the enormous denominator of ballots cast.

There is an additional advantage to framing it this way: you don't need to claim that voter mistakes change election winners. "Substantive" can mean that they occur in numbers large enough to matter electorally, particularly in close races. Whether they actually reverse a particular result is a separate empirical question.

So I would formulate your actual conclusion something like:

Even after subtracting voters who intentionally leave a contest blank, ordinary mistakes caused by ballot design, voting technology, instructions, or simple voter error appear to misrecord or lose far more votes in a typical major election than intentional voter fraud does. The former is routinely measurable in tens or hundreds of thousands nationally; documented intentional fraud is sufficiently rare that investigations generally measure it in isolated cases rather than comparable fractions of the electorate.

There is one distinction I'd preserve carefully. “People misunderstanding the voting method” is slightly too narrow. The empirical literature supports voter/interface error: confusing instructions, poor layout, equipment characteristics, accidental overvotes, failure to notice another page/column, etc. Some of those are misunderstandings; others are essentially usability failures. That's actually better for your underlying argument because you're talking about the mundane ways in which an election can imperfectly translate voter intent into recorded votes.

And the killer comparison isn't even necessarily error versus fraud totals. It's the difference in base rates. If an error mechanism affects only one voter in a thousand, that's 150,000 errors in a 150-million-voter election. For fraud to equal it, intentional fraudulent voting would also have to occur at roughly one ballot per thousand. Nothing in the fraud literature we've been looking at suggests anything remotely approaching a 0.1% national fraud rate.

That's the piece I'd want nailed down next: find the lowest empirically defensible post-abstention error rate rather than a central estimate. If even the lower bound dwarfs credible fraud estimates, the argument becomes much harder to attack because we have deliberately stacked the assumptions against the conclusion.

G: Yes. That is the comparison, and I think your formulation gets closer to the rhetorically accurate version than my attempts to manufacture a precise national estimate.

Taking the numbers we've established:

Heritage: ~1,500 documented fraud entries over ~25 years.

Now give the fraud side an intentionally ridiculous advantage and pretend every single Heritage entry involved 100 fraudulent votes:

1,500 × 100 = 150,000 fraudulent votes over 25 years.

Compare that with the deliberately low-end interface/error assumption we were using: 0.1% of ~150 million voters = ~150,000 votes in one presidential election.

So yes, the striking formulation is essentially:

You could pretend that every election-fraud case Heritage has documented for the last 25 years involved 100 fraudulent votes, and you'd only get about 150,000 votes. That's roughly the number you'd expect to lose or misrecord from ordinary voter/ballot-interface errors in a single presidential election even using an extremely conservative error rate.

And the 100-votes-per-Heritage-case assumption is deliberately absurd in fraud's favor. We already know many Heritage entries involve one or a handful of ballots, some involve no fraudulent ballots at all, and multiple entries can represent defendants in the same scheme. There certainly are larger cases, but an average of 100 fraudulent votes across ~1,500 entries would imply approximately 150,000 documented fraudulent ballots. If Heritage's own database actually established anything remotely like that, it would be an extraordinarily prominent statistic rather than something we have to manufacture by multiplying its case count by 100.

There's also a useful rhetorical distinction in the time periods:

25 years of documented fraud × an enormously inflated 100 votes per case ≈ one year's extremely conservative estimate of mundane voting error.

That's much cleaner than “fraud is X times more common” or “there have been exactly Y fraudulent votes.” Neither of those stronger claims is necessary.

The qualifier I'd keep is “presidential election” rather than “federal election year.” In a midterm, turnout is lower, and if we start counting House/Senate contests there are different residual rates and multiple contests. Presidential elections give you the cleanest apples-to-apples denominator: one presidential choice per participating voter.

So your intuition is right. We're not trying to determine the sixth decimal place of either phenomenon. We're establishing that they're on radically different scales. Giving the fraud side an implausibly generous multiplier still only gets twenty-five years of it into the same neighborhood as a deliberately conservative estimate of ordinary voting error from one presidential election.

G: Correct. That's an important additional asymmetry.

The Heritage database is not limited to presidential elections—or even federal elections. It aggregates election-fraud cases across federal, state, and local elections and includes things such as primaries and ballot/petition-related offenses.

So your comparison is actually more conservative than it first appears:

Take roughly 25 years of Heritage's documented election-fraud cases across elections at every level. Pretend every one involved 100 fraudulent votes. That produces ~150,000 votes—roughly comparable to an extremely low estimate of ordinary voter/interface error in the presidential contest alone in one high-turnout presidential election.

That's the rhetorically important apples-to-oranges distinction: the fraud side is being given 25 years, many kinds of elections, and an enormous hypothetical 100-vote multiplier; the error side gets one contest in one election.

I'd definitely preserve that when you use the comparison.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes—with one terminology correction.

A 0.5% rate is very defensible as a deliberately lowball estimate of presidential votes lost/misrecorded through voter + interface/technology error, but I would not call it a 0.5% “residual vote rate.” Actual measured residual rates have generally been higher, and residual votes also include intentional abstentions.

The Caltech/MIT Voting Technology Project found technology-related differences alone of up to roughly 1.5 percentage points, and estimated that replacing punch cards with optical scanning in 2000 would have resulted in roughly 500,000 additional presidential votes being successfully recorded. citeturn0search3turn0search4 Later research also found that jurisdictions changing voting technology experienced corresponding changes in residual-vote rates. citeturn0search36

So for rhetorical accuracy, I would be comfortable saying:

“Let's lowball ordinary voter/ballot-system error at just half of one percent—one vote in 200.”

That's conservative enough that you're not claiming 0.5% was independently measured as the exact error rate in every election. You're using it as a deliberately favorable assumption for the opposing comparison, supported by empirical rates comfortably large enough to make it reasonable.

Two good primary/academic references:

urlCaltech/MIT — Residual Votes Attributable to Technologyhttps://vote.caltech.edu/reports/5

urlMIT Election Lab — Residual Vote Rate data and definitionhttps://electionlab.mit.edu/landscapes/data

G: Yes. That reasoning is sound for the rhetorical purpose you're describing.

If the observed presidential residual rate is around 1%, and roughly 30% of residuals are estimated to be intentional abstentions, that gives ~0.7% potentially attributable to error. Dropping that again to 0.5% deliberately throws away another ~0.2 percentage points—almost 30% of the remaining estimate—to accommodate uncertainty, other causes, and imperfections in the decomposition.

So 0.5% isn't you selecting the high end to make the argument work. It's a deliberately conservative working assumption that remains comfortably inside the empirical evidence.

And that's probably the strongest rhetorical position: you don't need to establish whether the “true” number is 0.63%, 0.72%, etc. Your argument survives after intentionally giving away that uncertainty.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Using the official presidential vote totals compiled from FEC returns, and taking 0.5% (1 in 200) of each: citeturn0search4

Election	Presidential votes	0.5%
2004	122,295,345	611,477
2008	131,313,820	656,569
2012	129,085,410	645,427
2016	136,669,276	683,346
2020	158,383,403	791,917
2024	155,238,302	776,192
Total	832,985,556	4,164,928

So your deliberately lowball 0.5% assumption produces about 4.16 million potentially erroneous/misrecorded presidential votes since 2000, excluding the 2000 election itself.

And even the smallest single-election number here—~611,000 in 2004—is four times the hypothetical 150,000 we got by generously pretending every one of Heritage's ~1,500 fraud cases involved 100 votes.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes. I checked specifically for that distinction, including today's developments, and you do not have two separate piles of fraud numbers that need to be added together.

The Heritage Foundation's overall Election Fraud Database is the superset. Mail/absentee-ballot fraud is one of the fraud types contained within it. Heritage actually labels individual records "FRAUDULENT USE OF ABSENTEE BALLOTS"; for example, its records include single-ballot cases as well as larger absentee-ballot schemes. citeturn0search2turn0search3

There's even a useful numerical cross-check. Ballotpedia recently analyzed Heritage's database and found 385 Heritage cases involving absentee/mail-in voter fraud between 1993 and 2025, versus Heritage's roughly 1,600 total database entries. So those ~385 are inside the ~1,600, not additional to them. citeturn0news26turn0search14

Therefore, for our earlier calculation:

~1,500 Heritage cases since 2000 = all categories combined, including mail-ballot fraud.

You absolutely should not take ~1,500 and then add hundreds of mail-ballot cases to it.

And I think I see why this occurred to you today. The Trump administration has gone back to the Supreme Court seeking to implement its new USPS restrictions on mail voting before the 2026 midterms. The proposed system would require states to provide mail-voter information and use specified ballot-envelope/barcode procedures; the litigation concerns whether the federal government can impose those requirements and whether the system can actually be implemented this close to the election. citeturn0news34turn0news58

So if today's discussion is making it sound as though “mail-ballot fraud” is some enormous additional category sitting outside the voter-fraud numbers we've been discussing, no—it isn't. The Heritage figures we've been using already encompass that category.

navlistToday's mail-ballot developmentsturn0news34,turn0news58

I can monitor the Supreme Court fight and flag the next substantive ruling if you want.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes — that is essentially the structural criticism, and today's reporting makes one part of it considerably less hypothetical than when we started this discussion.

The one thing I'd modify is “386 cases where such a thing has maybe occurred.” Heritage's absentee/mail category consists of proven election-fraud cases, not merely suspicions, and some cases involve multiple ballots while others overlap as defendants in one scheme. So I'd say “roughly 385 documented absentee/mail-ballot fraud cases in Heritage's database over roughly three decades.” Heritage itself describes the larger database as a sampling, so you shouldn't imply those are necessarily every case that occurred. citeturn0search12turn0search14turn0search15

But your asymmetry survives those qualifications extremely well. Today's actual proposal would require states to submit mail-voter lists to USPS and use specified envelopes with unique barcodes; noncompliant ballots could be rejected. And this isn't merely critics imagining that a complicated new system might malfunction: USPS is still developing the system as voting begins, and an internal whistleblower has described the implementation as rushed and error-prone. citeturn0news24turn0news32

More importantly, we already have the failure mode you're describing occurring in the real world. Election officials are reporting increased mail-ballot rejection associated with postal/postmark problems, while North Carolina began sending ballots today, September 4, amid continuing uncertainty over what rules will ultimately govern them. citeturn0news27

So I think the strongest version of your argument is:

The policy is attempting to reduce an empirically rare form of intentional wrongdoing by adding additional failure points to a process used by millions of legitimate voters. That trade could make sense if the added procedure prevented substantially more bad ballots than good ballots it caused to be rejected. But given the enormous difference in the underlying base rates, even a very small procedural-error rate can overwhelm the amount of fraud the procedure is intended to prevent.

That's the mathematical heart of it. You don't even need to establish malign intent to make the criticism. It's an engineering/base-rate objection: when protecting against a rare event, a safeguard with a false-positive or failure rate considerably greater than the event rate can make the system less accurate overall.

The second issue you're identifying is separate but mutually reinforcing: introducing an inadequately tested procedure likely to cause delays or disputes is especially dangerous after years of rhetoric treating delayed vote counts as evidence of wrongdoing. I would keep that as a second argument rather than bake it into the numerical one, because the first argument stands without having to establish anyone's motives.

navlistToday's mail-voting developmentsturn0news24,turn0news27,turn0news32

G: Yes. I understand the distinction you're making now. Rhetorical compression isn't the same thing as factual sloppiness. A sentence can deliberately discard qualifications when those qualifications don't materially change the proposition—and sometimes including them actually communicates the underlying reality less accurately by burying the scale comparison.

Your “386th case” construction is a good example. Literally, a Heritage case isn't necessarily one ballot. But rhetorically you're deliberately granting the hypothetical new case significance while emphasizing the observed frequency of the phenomenon. Stopping mid-argument to explain defendant duplication, multiple-ballot cases, database incompleteness, etc. would obscure rather than illuminate the point you're making.

And the Obamacare analogy is unusually clean for the implementation objection. You can separate the policy question entirely:

Forget whether this is a good idea. You don't deploy version 1.0 of a new nationwide voting system during an election. Obamacare should have taught everybody that lesson.

That invokes entity["organization","HealthCare.gov","U.S. federal health insurance marketplace website"]'s disastrous 2013 launch without requiring agreement about whether the Affordable Care Act itself was good policy. Someone can believe the ACA was excellent legislation and still acknowledge that its initial technical rollout was a mess. Conversely, somebody who hated Obamacare already accepts the example.

There's an even simpler engineering principle underneath both:

A system can be conceptually sound and still be operationally unready.

With healthcare enrollment, a broken version 1.0 could eventually be repaired. With an election, you've got a particularly nasty constraint: you can't patch Election Day afterward. And if the new procedure itself determines whether otherwise legitimate ballots are accepted, its failure modes aren't peripheral—they directly affect the thing the system is supposed to protect.

I've also got your rhetorical preference: unless precision itself is material to the argument, I'll favor short, substantively accurate compression over qualification-heavy formulations, while flagging something if the compression actually changes the conclusion rather than merely shaving off precision.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes. I looked specifically for examples where the lesson isn't merely “version 1 had bugs”, but “known engineering/implementation practice said this needed testing before exposure to real-world consequences.” A few American examples are particularly good.

HealthCare.gov, 2013 — probably your cleanest political analogy. GAO later found that the system was not fully tested before launch, testing was inconsistent, known coding errors hadn't been adequately corrected, capacity planning was inadequate, and CMS launched without verification that the system met performance requirements. This wasn't hindsight discovering an unforeseeable problem; ordinary systems-development practices already called for doing this work. citeturn0search0turn0search9
urlGAO investigation of the HealthCare.gov launchturn0search0
Knight Capital trading disaster, 2012 — fantastic nonpartisan/private-sector analogy. New trading software was deployed incorrectly to eight servers; one server retained obsolete code. There was no required second-person review, inadequate deployment/testing procedures, and inadequate safeguards. When the market opened, the system generated more than 4 million erroneous orders in 45 minutes and Knight lost over $460 million. The SEC's conclusion was essentially your principle: when failure can have enormous consequences, you test the components and build safeguards before going live. citeturn0search1turn0search48
urlSEC account of the Knight Capital failureturn0search1
Denver International Airport automated baggage system, 1990s — excellent physical-system analogy. Denver committed to an extremely ambitious airport-wide automated baggage system. Testing produced misrouted bags, bags falling from carts and system jams; GAO reported that it hadn't successfully passed the tests necessary for certification. The airport opening was repeatedly delayed and Denver ultimately installed a conventional backup baggage system. In other words: they discovered during testing precisely why you don't make the first full-scale implementation the production environment. citeturn0search3turn0search4
urlGAO report on Denver's baggage-system failureturn0search3
FBI Virtual Case File, early 2000s — good government/mission-critical example. The FBI spent years and roughly $170 million trying to replace its antiquated case-management system, but poor requirements, integration planning, project management and oversight eventually made the system effectively undeployable. The FBI abandoned it and started over with Sentinel. GAO's subsequent work explicitly emphasized disciplined acquisition and testing practices because of what happened with VCF. citeturn0search2turn0search14

For what you're arguing, though, I'd use HealthCare.gov + Knight Capital. They make the principle almost impossible to partisanize:

We already know what happens when you make the production environment your beta test. HealthCare.gov did it and crashed; Knight Capital did it and lost $460 million in 45 minutes. Whatever you think of the underlying policy, an election is an exceptionally stupid place to try version 1.0.

And there's an important refinement to your earlier wording: “stress testing” is only part of the objection. What you're really asking for is validation before production—ordinary functional testing, integration testing, load/stress testing, controlled pilots, and enough elapsed time to discover failures and correct them. A system can pass a stress test and still reject legitimate ballots because somebody didn't anticipate an interaction between two perfectly functioning components. That's exactly why you don't want the first nationwide deployment to double as the integration test.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Ah, yes. You want cultural shorthand, not case studies requiring a paragraph before anybody recognizes the example.

In that case, there are several much better candidates:

Ford Pinto — recognizable, but actually not quite your phenomenon. Ford did crash-test it and saw fuel-system failures before release. That's more “we knew it was dangerous and shipped it anyway” than “we made production the beta test.” citeturn0search13turn0search45
Space Shuttle Challenger — extremely recognizable and closer to the principle, although again the problem wasn't absence of all testing. The Rogers Commission found inadequate qualification testing, unresolved O-ring problems, and failure to test adequately under actual flight conditions. The cultural shorthand is basically: “You don't find out whether the O-rings work in the cold by launching Challenger.” citeturn0search0turn0search6
Samsung Galaxy Note 7 — probably the best consumer-product example from recent memory. Phones went into consumers' hands, batteries started catching fire, Samsung recalled about a million—and then some of the replacement phones also overheated, forcing an expanded recall of about 1.9 million devices. That's remarkably close to “version 1.0 shouldn't be tested on the public.” citeturn0search3turn0search4
HealthCare.gov — still probably the best software rollout analogy because practically everyone of political age remembers “the Obamacare website didn't work.”

But Challenger may give you the rhetorical form you're actually looking for:

“You don't test a new system by making the first real use the test. We learned that lesson with Challenger. You test the damn thing under the conditions it's actually going to face first.”

And there's an unusually good factual foundation underneath that compressed version: the Rogers Commission specifically concluded that the joint qualification-testing program was inadequate and that NASA/Thiokol had failed to develop and verify a replacement even after testing and flight experience exposed the deficiency. citeturn0search0

If you want something even more culturally ubiquitous than these, I can hunt specifically through cars, bridges/buildings, consumer products, military hardware, and government programs from ~1900 onward for five or six “oh, that disaster” examples where insufficient predeployment testing really was central, rather than merely adjacent.