Sometime this summer, the Department of Homeland Security made a small administrative decision. It lowered a number. The target accuracy rate for a database called SAVE — the system the federal government uses to check whether a registered voter is a citizen — was revised from 99 percent down to 97 percent.
Two percentage points. The kind of change that gets disclosed to a reporter and absorbed into a single afternoon’s news cycle. No press conference. No rule. Just a target adjusted, confirmed by a U.S. Citizenship and Immigration Services (USCIS) spokesman to the Washington Times in May.
Here is what that number means when you do the arithmetic. There are roughly 200 million registered voters in the United States. A 3 percent error rate, applied across that population, is not an abstraction. It is approximately six million people. And under a rule the Postal Service proposed on May 29, the people that database cannot cleanly confirm as citizens are the people who will not receive a mail ballot this November — with no warning, no notice, and no way to fix it inside the window that matters.
The most alarming part is not the 3 percent. It is what happens to the people inside it. When states have actually run their voters against SAVE and then checked the results by hand, most of the voters the database flagged as potential noncitizens turned out to be eligible American citizens — in the most rigorously documented cases, between 64 and 87 percent; in the full range of documented cases, as low as 14 percent in a single county, as high as 87 percent statewide. Not noncitizens correctly caught. Citizens wrongly accused.
This is the story of how a rounding error becomes disenfranchisement — and how the architecture was built so that no court will look at it until after the ballots aren’t mailed.
Before we continue, one quick ask.
A single raid is a headline. The pipeline behind it is a story nobody’s funding.
At The RAMM I connect the dots that don’t fit in one article: the federal contracts, the 287(g) deals, the quiet real-estate moves, the county-by-county buildout of detention capacity.
The receipts:
4,776+ sourced events at CaptureCascade.org
1,988 counties with signals of detention expansion at detention-pipeline
129 community fights over detention capacity, tracked
This is self-funded. Paid subscriptions are the only thing that keeps it alive.
What SAVE is, and what it was never built to do
SAVE stands for Systematic Alien Verification for Entitlements. The name tells you what it was for: verifying immigration status so that government agencies could decide who qualifies for benefits like Medicaid, SNAP, and a driver’s license. It was built to answer one question — is this person a noncitizen who is nevertheless lawfully present? — for people who had already raised their hands and applied for something.
It was not built to verify U.S. citizenship. That is a structurally different question, and the difference is the entire problem.
SAVE contains Department of Homeland Security (DHS) immigration records — green cards, visa filings, A-numbers — plus Social Security citizenship and work-authorization data, and, since a 2025 overhaul, motor vehicle records. What it does not contain is a complete, current census of American citizens. It has no comprehensive record of citizens born in the United States. It has stale, incomplete records on the people most likely to be queried and flagged: naturalized citizens, who appear in the system because they were once noncitizens, and whose files may never have been updated to reflect that they took the oath.
A naturalized citizen is, by definition, a person SAVE once knew as a noncitizen. When you run the entire voter file against a database like that and ask it to certify citizenship, the false positives are not bugs. They are the predictable output of asking a benefits-eligibility tool to do citizenship verification it was never designed to perform. You cannot patch this with better data hygiene. The category mismatch is baked into what SAVE is.
In April 2025, DHS, USCIS, and the Department of Government Efficiency (DOGE) announced a “comprehensive overhaul” of SAVE to enable exactly this — mass voter verification. The overhaul added motor-vehicle data and expanded the system’s query volume from a few hundred thousand benefit checks a year to tens of millions of voter-registration checks. This is the move worth naming clearly. It was not a scale expansion of an existing function. It was a category expansion: the infrastructure of immigration enforcement, repurposed as the infrastructure of voter verification. The database name didn’t change. The population it can harm did.
The arithmetic, laid out so you can check it
Let me make the math transparent, because the math is the argument. Every figure here traces to a documented source, and I want you to be able to use these numbers yourself — including in a comment to the Postal Service, which I will get to.
Start with two different error rates, because conflating them is how this story gets reported badly.
Rate one is the inconclusive rate — the share of voter queries for which SAVE simply cannot return a definitive answer. The most precise figure available comes from the Bipartisan Policy Center, which in August 2025 analyzed SAVE’s performance on actual voter-verification queries and found a 3.12 percent inconclusive-response rate. the Bipartisan Policy Center (BPC) itself noted that “while 3 percent might seem small, it represents potentially millions of voters whom election offices would need to contact or investigate.”
That 3.12 percent maps almost exactly onto DHS’s own downgraded 97 percent accuracy target. The two figures are not the same measurement — an inconclusive result (SAVE declines to answer) is conceptually distinct from an error (SAVE returns a wrong answer) — but they converge on the same population: voters the system cannot cleanly confirm. That convergence is what matters operationally, because the gate keys on inconclusive results, not on formal errors.
Now apply it. Using ~200 million registered voters as a conservative midpoint between the Census Bureau’s survey figure (~174 million) and the Election Assistance Commission’s administrative figure (~211 million):
200,000,000 × 0.0312 = 6,240,000 inconclusive results
At the administration’s stated 97 percent target, the figure is 6 million. At the 98 percent accuracy DHS acknowledged it actually achieved in 2025, it is 4 million. So the honest range for voters SAVE cannot cleanly confirm is four to six million.
That is rate one. By itself it would be a manageable administrative headache — a lot of cases needing a second look.
Rate two is the false-positive rate — of the voters who do get flagged, how many are actually eligible citizens. This is the number that turns an administrative headache into disenfranchisement, and it is documented not in projection but in the field, in states that already ran this experiment:
Iowa: The Secretary of State flagged 2,176 potential noncitizens through SAVE. After investigation, the confirmed count dropped to 277. That is an 87 percent false-positive rate — 1,899 eligible citizens wrongly flagged.
Alabama: Secretary of State Wes Allen moved to purge 3,251 registered voters identified through SAVE, 84 days before the 2024 election. A federal judge blocked it; the court record showed roughly 2,074 of the flagged voters — 64 percent — were confirmed eligible citizens.
St. Louis County, Missouri: Of 691 flagged voters, 81 percent of the original data was wrong once passports were checked, leaving 133 genuine noncitizen registrants out of the original 691.
Boone County, Missouri: Of 74 flagged voters, more than half were citizens. One flagged voter’s registration paperwork bore the initials of the very staff member who had helped them register — at their own naturalization ceremony.
Denton County, Texas: 84 flagged; at least 12 immediately proved citizenship — a documented floor of 14 percent, which the county official acknowledged was “probably higher.”
So: between 14 percent (the lowest documented floor) and 87 percent (Iowa) of flagged voters are eligible citizens, with most documented cases clustering between 60 and 87 percent.
Multiply the two rates. A note on the method: the inconclusive rate (3.12%) comes from BPC’s national analysis of SAVE voter-verification queries; the false-positive rates come from state-level samples of voters who appeared on flagged lists. These are different sampling frames, so the multiplication is an order-of-magnitude estimate, not a precise count — but that is exactly why the floor matters. Take the four-to-six million flagged and apply even a conservative midpoint false-positive rate of 60 percent:
4,000,000 × 0.60 = 2,400,000 (low estimate) 6,000,000 × 0.65 = 3,900,000 (central estimate)
The central estimate is between three and four million eligible American citizens flagged as potential noncitizens. The high end, using the documented 87 percent rate, reaches above five million. And the absolute floor — using only Denton County’s 14 percent, the lowest false-positive rate anyone has recorded — is still 560,000 citizens.
The floor is more than half a million wrongly flagged voters. The central estimate is in the millions. There is no version of this arithmetic that produces a small number.
A note on what the record does not contain, because the absences are load-bearing too. There is no DHS Inspector General audit of SAVE’s accuracy for voter verification. There is no Government Accountability Office (GAO) audit of SAVE since 2017 — and the 2017 report (GAO-17-204) covered benefit eligibility, not citizenship checks on the voter file. The full USCIS submission to Congress with the 98 percent actual-accuracy figure has not been publicly released; it currently exists in the public record only as relayed by the Washington Times. No federal watchdog has independently audited the accuracy of the tool now being pointed at the electorate. We are running it live.
The one-way gate
For most of SAVE’s history, a false positive was survivable, because the system that used it included a way to argue back. If a benefits agency flagged you, you got a notice. You could produce your naturalization certificate or your passport. The agency was required to complete additional verification before issuing a final denial. The design assumed error and built in a correction step.
Under the Postal Service’s proposed rule — Federal Register Docket 2026-10968, published May 29 to implement Executive Order 14399 — that correction step does not exist inside the window that decides whether you vote. Here is the operational sequence:
DHS runs the voter files against SAVE and builds State Citizenship Lists of confirmed citizens.
Voters with inconclusive SAVE results are not included on those lists.
States transmit their mail-ballot lists, derived from the federal lists, to USPS at least 30 days before ballots are mailed.
USPS mails ballots only to voters on the list.
USPS has expressly stated it “would not verify whether individuals should or should not be included” — because, in its framing, the verification already happened upstream. The Postal Service is the delivery point, not the review point.
Step two is the gate. A voter with an inconclusive SAVE result is not on the list. They receive no notification that they are not on the list. There is no mechanism at the postal tier to challenge it. And the timeline gives them no room to discover the problem and fix it: most states begin mailing ballots about 60 days before Election Day — early September — and the list is locked 30 days before that.
The voter does not know in advance that no ballot is coming. By the time they realize a ballot never arrived, the window has closed. A ballot not mailed is a vote not cast, and there is no downstream correction inside the mail-ballot process.
The proposed rule also requires a unique barcode on each ballot envelope, tied to the federal list entry, enabling per-ballot tracking. That detail matters for one reason: it tells you the non-delivery is not an accident the system is straining to prevent. The infrastructure to confirm that the right people are being excluded is being built into the envelope. The gate is engineered, not improvised.
This is the difference between the old SAVE and the new one. Under benefits, a false positive cost you a delayed Medicaid approval and a phone call. Under Docket 2026-10968, a false positive costs you your vote, silently, with no one obligated to tell you it happened.
Why no court will stop it in time
You might assume that a mechanism capable of failing to mail ballots to millions of citizens would be the kind of thing a federal court would examine before it goes live. It will not — and the reason is a quiet masterpiece of timing.
On May 28, 2026, U.S. District Judge Carl J. Nichols (D.D.C., a Trump appointee) denied a motion for preliminary injunction in DSCC v. Trump, the consolidated challenge to EO 14399 brought by the Democratic Senatorial Campaign Committee alongside the NAACP, the League of United Latin American Citizens (LULAC), Common Cause, Black Voters Matter, and the Lawyers’ Committee for Civil Rights Under Law. He did not rule on whether the executive branch can set the rules for federal elections — a question Article I of the Constitution assigns to state legislatures and Congress. He ruled on ripeness — the legal doctrine that a court will not decide a challenge until the harm is concrete rather than anticipated.
His logic: no final rule existed yet, so no concrete harm had crystallized. He wrote that “the Postal Service may ultimately issue a final rule that directly affects Plaintiffs,” and that the plaintiffs “may renew their motions if and when those future actions occur.” Until then, he found the challenge “largely speculative” (docket 1:26-cv-01114, Mem. Op., May 28, 2026). Reasonable enough, in the abstract. The harm hadn’t happened yet.
The day after Nichols ruled — May 29 — the Postal Service published its proposed rule. The procedural obstacle was cleared on a Wednesday; the executive branch walked through the opening on a Thursday. The sequencing is documented, in the public record, one day apart.
Now follow the calendar forward, because the calendar is the weapon:
Comment period closes: July 2, 2026.
Final rule expected: ~July 29, 2026 (the EO’s 120-day mandate from its March 31 signing).
States begin mailing ballots: early September 2026.
Window between final rule and ballot transmission: roughly 30 days.
A party seeking to stop the final rule would have to file, brief, argue, and win — all within about 30 days of the rule’s publication, before ballots go out — while the government argues, again, that the harm is still speculative until ballots are actually refused. The same ripeness logic that foreclosed relief in May will be raised again in August. And if a court holds that it needs evidence of actually-refused ballots before it will find a cognizable injury, then by the time the evidence exists, the ballots have already not been mailed. The harm and the proof of harm arrive at the same moment, which is the moment it is too late.
This is the trap, and it is elegant. The ripeness doctrine is being applied correctly at every step. There is no rogue judge, no defied order, no obvious villainy in the opinion. The architecture simply sequences the harm to land after the final rule and before litigation can reach the merits. Every move is formally compliant with administrative law. That is not a softening of the problem. That is the problem. The timing is the weapon: by the time a court is allowed to rule on whether this is legal, the ballots have already not been mailed. The damage is finished before the law is allowed to look at it.
Who built it, and who can’t be found
There is a named conductor, and there is a void above him, and both are part of the finding.
The named official is Postmaster General David Steiner, who signed the proposed rule. According to Democracy Docket’s reporting, Steiner had “met with other senior Trump administration and DOJ officials to discuss ways to implement the order” before the rulemaking. That is not inferred coordination — it is a disclosed inter-agency steering process. The one-day gap between the Nichols ruling and the USPS notice is not a coincidence; it is what coordination looks like from the outside.
Above Steiner, the only other verifiably named coordinator is Commerce Secretary Howard Lutnick. EO 14399’s text directs DHS and USPS to “coordinate with the Secretary of Commerce in effectuating all relevant aspects” of the order — an instruction with no obvious statutory basis, since Commerce has no authority over either the Postal Service or elections. Senate Democrats flagged the anomaly in a June 1, 2026 letter and characterized “nonpublic reports that you and your office had a key role in drafting” the order. I want to be precise about what that is: a congressional characterization of a nonpublic claim, not independent confirmation that Lutnick drafted the EO. It is attributed here as exactly that and no more.
Above Lutnick, the record goes dark — and the darkness is itself the finding. The usual architects of this kind of policy — Stephen Miller, Russell Vought, James Sherk, the longtime voter-fraud advocates Hans von Spakovsky and Cleta Mitchell — are not documented in any primary or independent source as architects of EO 14399. The mechanisms in the order map onto von Spakovsky’s and the Heritage Foundation’s published advocacy with real precision: SAVE-based voter verification, federal-database matching, mail-ballot restrictions. DHS did brief Mitchell’s Election Integrity Network on SAVE implementation under a predecessor order in June 2025. The information flowing out of the administration to the advocacy network is documented. The flow back — the drafting credit — is not.
This is the deniable-proxy pattern, and it is worth understanding as a category. Procurement capture leaves a paper trail: ethics filings, SEC disclosures, contracting databases. Executive-order capture leaves almost none. The mechanisms are public; the attribution chain above the named implementing officials runs through deliberative-process privilege and channels designed to leave no record. The policy is fully documentable. The architects above the postmaster general are not. That asymmetry is not a gap in my reporting. It is the structural signature of how this kind of power now operates.
What this is, in one sentence
A regulatory mechanism produces an at-scale operational effect — millions of citizens unable to vote by mail — inside a judicial-review window that has been foreclosed by the correct application of ripeness doctrine. Every step is legal. The engineering is in the sequence.
SAVE was built for benefits. It was overhauled in 2025 to check tens of millions of voters against a database that cannot reliably tell a naturalized citizen from a noncitizen. Its accuracy target was lowered, quietly, to 97 percent. At the scale of the electorate, that produces millions of inconclusive flags, most of which fall on eligible citizens. Those citizens drop off the list that decides who gets a mail ballot. They are not told. And the one court that could have looked at the merits declined, on the ground that the harm had not happened yet — the day before the machine that produces the harm was switched on.
What you can do, before July 2
This is where the timing cuts both ways. The same narrow window that forecloses litigation is also, right now, open for something else: the public record.
The Postal Service’s proposed rule is in a public comment period that closes July 2, 2026. Anyone can submit a comment on Docket 2026-10968 through the Federal Register at federalregister.gov/documents/2026/06/02/2026-10968/ballot-mail-for-federal-elections. Click “Submit a Formal Comment” on that page; no account is required. You can also mail written comments to: USPS Headquarters Library, 475 L’Enfant Plaza SW, 11th Floor North, Washington, DC 20260. Comments become part of the official rulemaking record. They are what a future court reviews when it asks whether the agency considered the foreseeable consequences of its rule. An agency that finalizes a rule over a record full of documented, specific, sourced objections is in a materially weaker position than one that finalizes over silence.
The arithmetic in this piece is built to be used. The 3.12 percent inconclusive rate is the Bipartisan Policy Center’s, published. The false-positive range — from a documented floor of 14 percent in a single Texas county to 87 percent statewide in Iowa, with most documented cases clustering between 64 and 87 percent — is on record in Iowa, Alabama, Missouri, and Texas, in court filings and county data. The one-way structure of the gate — no notice, no challenge, no correction inside the mailing window — is in the text of the rule itself.
After July 2, this becomes history. Before July 2, it is a warning that can still be entered into evidence.
Three to four million eligible American citizens. A database that was never built to find them. A rule that turns its errors into silence. And a courthouse door that opens only after the ballots have already failed to arrive.
The comment period closes July 2. The ballots start mailing in September. The window between them is the whole story.
Sources
Primary timeline events (Capture Cascade): - Judge Carl Nichols declines to block EO 14399 mail-ballot provisions (May 28, 2026) - USPS issues proposed rule, Docket 2026-10968, implementing EO 14399 (May 29, 2026)
Government and primary documents: - U.S. Postal Service, Proposed Rule, Federal Register Docket 2026-10968 (published June 2, 2026) — comment period closes July 2, 2026 — federalregister.gov - Executive Order 14399, “Ensuring Citizenship Verification and Integrity in Federal Elections” (signed March 31, 2026) — Federal Register - GAO-17-204, “Immigration Status Verification for Benefits: Actions Needed to Improve Effectiveness and Oversight” (2017) — gao.gov - U.S. Census Bureau, Voting and Registration in the Election of November 2024 (CPS) — ~174M registered — census.gov - U.S. Election Assistance Commission, 2024 EAVS Report — ~211M active registered
Policy research: - Bipartisan Policy Center, “What Adding Motor Vehicle Data to USCIS’s SAVE System Means for Election Administration” (August 2025) — 3.12% inconclusive rate; Iowa case — bipartisanpolicy.org - Protect Democracy, “The SAVE Tool Explained” — Alabama 3,251 flagged / 2,074 eligible — protectdemocracy.org - Brennan Center for Justice, “Homeland Security’s ‘SAVE’ Program Exacerbates Risks to Voters” — brennancenter.org
Reporting: - Homeland Security lowers the accuracy rate for database used to spot voter fraud (Washington Times, May 15, 2026) — USCIS spokesman Zach Kahler; 97% target; 98% actual 2025 - SAVE tool keeps mistakenly flagging voters as noncitizens (Texas Tribune, Feb 13, 2026) — Denton County - “The Broken Database That Could Upend the 2026 Election” (Popular Information) — St. Louis County, MO - “Not Ready for Prime Time” (Techdirt / Alaska Beacon, Feb 26, 2026) — Boone County, MO - Federal judge blocks Alabama voter purge program (NBC News, Oct 2024) — Alabama - Postal Service moving forward with Trump’s attack on mail voting (Democracy Docket, May 29, 2026) — Steiner coordination; SAVE reliance - Federal judge refuses to block Trump order on mail ballots — for now (Votebeat, May 28, 2026) — Nichols ruling
Related RAMM coverage: - Trump’s Campaign to Decide Who Counts? (The RAMM, April 3, 2026) — the SAVE+USPS mail-ballot architecture and the birthright-citizenship layer of the same electorate-control system - Every Database Is an Immigration Database Now (The RAMM, April 1, 2026) — the pattern of which SAVE’s repurposing is the continuation - The Lutnick System (The RAMM, April 10, 2026) — Commerce Secretary Lutnick’s conflict architecture and the anomaly of Commerce’s coordination role in EO 14399