Executive Summary

Every figure here is as received on November 7, 2026 (KST). Counting is not over and no result has been certified. The raw data was downloaded directly from Ballotpedia's public table and from the public repository of the Redbud Consulting tracker, and computed here.

On the November 3, 2026 midterm ballot, questions about data centers went before voters in forty-one places across ten states. Four months earlier this was a one-town story. Monterey Park, California, became the first U.S. city to block a data center by popular vote in June 2026, with 88% voting yes. On November 3 the same question was printed on forty-one ballots at once.

Three days ago this blog described how those forty-one measures were assembled before the vote and closed by noting that no returns were in at the time of writing. This article is the next page of that sentence. The substance of the opposition is covered in an earlier survey analysis, and the power problem driving it in the power wall report, so neither is repeated here. This section does one thing. It sets down what happened, all of it, without a verdict.

By state the distribution leans hard one way. Ohio alone accounts for twenty-one measures, more than half the total, followed by seven in Illinois, four in Maine, three in Kansas, and one each in California, Florida, North Dakota, New Jersey, Nevada, and Wisconsin. Together the forty-one measures drew 252,709 votes. Not one of them went before a whole state. These were city, town, township, and county questions, and in the smallest place a few hundred people settled it.

Three different instruments also sit under the single phrase "ballot measure." Twenty-five were citizen initiatives, put on the ballot by residents gathering signatures; eleven were referrals placed there by a city or its council; and five were referendums asking whether to overturn a decision already made. Nine of the forty-one are advisory questions with no legal force. This distinction returns later, at the point where the denominators part ways.

Laid out along a single line, the shape of the distribution becomes visible. The figure below places one dot for each of the forty-one measures at the share of the vote won by the side that restricts data centers. On a restricting measure that is the yes share; on an allowing measure it is the no share. One dot is one measure, and the farther right it sits, the more heavily the restricting side won in that jurisdiction.

Measures of quite different kinds sit under that one word. Harwood and Grove City, Ohio, were reacting to a specific project that already had a name. Measure DB in Alhambra, California, by contrast, is a citywide blanket ban with no disclosed project behind it, and even then its job was to lock a ban the city council had already adopted in June 2026 into the city charter by popular vote. Had it failed, the ban would have remained. Maine's four are not one kind either. Three are moratoriums of a fixed length, and the fourth, North Berwick, is a size threshold barring facilities above 25,000 square feet.

Below is the full forty-one-row table. The "What YES means" column records which side a yes vote falls on for that measure, and the rightmost column, "Restricting %," holds the values plotted in the figure above. Neither column was printed on any ballot; whoever did the tallying added them. Section 3 is about that column.

Source: the 41 rows of the November 3 table in Ballotpedia's "Data center-related ballot measures, 2026," transcribed and checked here (as received 2026-11-07 KST). The Type column keeps Ballotpedia's own classes: 25 initiatives, 11 referrals, 5 referendums. Alhambra, California, was still counting mail ballots at this point, and Los Angeles County's official certification date is November 27, 2026. Boulder City, Nevada, has a settled outcome of rejection, but total votes differ from one source to the next, so this table uses Ballotpedia's figures.

Gather up the figures the headlines carried about that day and at least ten of them are in circulation. 40 to 1, 46 to 1, 39 of 40, 46 of 47, 38 of 39, 38 to 1, at least 44, at least 38, 33, and 27 to 14. Every one of them points at the same day and none of them is wrong. The table below puts all ten in one place, with what each line actually counted in the right-hand column. This section collects what circulated outside; the figures we rebuilt ourselves are kept separate, in section 5.

All ten lines point at the same day. Each source is listed with its date in the references.

The ten lines part ways at four points and no more. Which list was used, whether a direction label was applied before counting, which moment in the returns was taken as the snapshot, and whether the count covers November 3 alone or the whole year. The first two belong to whoever builds the data, so sections 3 and 4 take them one at a time. The last two are conditions on how much of an already-built file to draw from. The filtering condition comes first.

Of the four, the snapshot is the plainest. Counting does not finish all at once, and the tracker records how far each row has come in one of four values. final for a count that is certified, all_precincts for every precinct reporting, counting for a tally still under way, and no_results for a row with nothing in it yet. Where the line for "a result" falls changes what comes out of one file. We filtered the same 47 rows three ways, working from the copy received on November 7, 2026, Korea time.

One file, one day, three headlines. The only thing that changed is a line of filtering. Source: the 47 November 3 rows of the Redbud tracker's results file, filtered on the reporting-stage column and computed here (as received 2026-11-07 KST).

Reading the line "20 to 0 if you count only certified rows" as a sign that the tracker padded its numbers gets it backwards. We can produce those three figures at all because the tracker writes a reporting stage on every row and published the whole file. Without that column only "46 to 1" would be visible, and there would be no way to ask which snapshot produced it. The structure in which the party that publishes its snapshot is the one that draws suspicion for it turns up once more further on.

Count only the Result column of the full table in section 1 and it comes to 27 passed, 14 defeated, a pass rate of 66%. The scorecard Ballotpedia builds from that same table, though, records 46 for the side limiting data center development and 1 for the side allowing it. Take November 3 alone and it is 40 to 1. The two figures are not drawn from two different tables. They come from the same forty-one rows, with a single recoding step in between.

The recoding goes like this. A ballot box counts yes votes and no votes on each measure, which is how every row gets a pass or a fail written against it. The record stops there. Then a person adds one more column to each row. On this measure, is a yes vote the side that blocks a data center, or the side that lets one go up? With that column in place, "defeated" splits into two meanings. A blocking measure defeated means the data center side won; an enabling measure defeated means the restricting side won.

In figures: of the forty-one measures, 28 were restricting and 13 were allowing. Among the 28 restricting measures, 27 passed and only Harwood failed. Among the 13 allowing measures, not one passed. That is how the restricting side's result becomes 27 plus 13, or 40, and the development side's result becomes 1.

Counted as passes and failures it reads "fourteen lost"; with direction added it becomes "only one lost." Both come out of the same table.

Neither of the two is wrong, and neither of them is hidden. Ballotpedia states the recoding in a sentence on its own page.

Two routes, folded into one count of 46: approving a restriction or a requirement, and rejecting a measure that would have eased development along. The second tracker, Redbud Consulting, puts the same operation more bluntly in its methodology document. Every question is mapped to Restrict (an outcome against data centers) or Allow (an outcome in favor), and this is what the document calls the column that holds the mapping.

One switch, and with it set correctly the rest falls into place. Turned around, the sentence says that a nationwide headline is held up by one column of a CSV file, and that a person filled that column in by hand. The risk does not sit with the person who filled it in. It sits with a reader who counts the Result column alone and carries the figure over as "data centers lost."

How much judgment that column asks for shows up in two rows of the same table. The measures in Monroe Township and Sprigg Township, Ohio, were about adopting zoning. The words "data center" never appear in the ballot language. Both registries put the pair on the data center list all the same, because the local fight was over a data center on a former power plant site, and a township gains regulatory authority only once zoning exists. The label follows what a measure is aimed at rather than the words printed on the ballot, and a person made that call, not a machine.

Academic work does have a yardstick for the quality of a hand-assigned label. Content analysis gives the same material to two or more coders separately and computes their agreement as intercoder reliability. With a single coder the measure cannot be computed at all. The direction label is a column one coder filled in, so no reliability figure exists for this event. That is not a fault in the tracker. Putting a second coder on volunteer-scale work running through election night would be the unusual choice. One thing is worth setting down. The headline that traveled the country has never had that column measured.

Coding rules producing numbers is not peculiar to this event. The U.S. count of people killed in police shootings makes a useful control. The Washington Post has run its own tally, Fatal Force, since 2015, and its public repository notes that the fatal shootings it has documented since 2015 run at more than twice the annual average in federal records, and that the FBI tracked about a third of all fatal shootings as of 2021. An intuition inverts here. The Post's criterion is narrower. It counts only shootings by officers in the line of duty. The narrower criterion produces the larger number because the FBI count depends on police departments submitting voluntarily. The guess that a loose definition makes a big number comes apart right here.

Conflict research has a sharper case. ACLED and UCDP, the two leading conflict-event datasets, record event counts for organized violence that run four to five times apart over the same period and the same conflict. A peer-reviewed paper from 2025 locates the cause not in the two looking at different sources but in their coding the same source sentences differently. For Syria in 2017, more than 90% of the gap between their fatality counts traces to one auxiliary rule, the one converting events whose death toll was reported vaguely into a fixed value set by the codebook. One codebook pins that value at 2, the other at 10. The paper sits behind a subscription, so we could not open the original, and it is paraphrased here rather than quoted.

Two organizations built the list of measures for that day. One is Ballotpedia, which gathers U.S. election information encyclopedia-style. The other is "Data Centers on the Ballot," a tracker opened ahead of the election by Redbud Consulting, a polling and qualitative research firm. For the same day the first list records 41 measures and the second records 47. Both also describe the reach of their data as "13 states," and the two thirteens are not the same thirteen.

We reconciled 41 against 47 first. Matched row by row, the containment is clean. Redbud's 47 holds all 41 of Ballotpedia's rows, and there is not one measure in Ballotpedia missing from Redbud. The difference is exactly six measures, and every one of them is restricting. Four are town meeting warrant articles from small Maine towns, namely Brownfield, Eliot, Freedom, and Sabattus, and the other two are Americus, Kansas, and Belvidere, New Jersey.

One explanation has to be cleared out of the way first. A summary went around saying the denominators diverge over whether advisory questions get counted, or whether rezoning referendums do. Opening the raw files does not bear that out. Ballotpedia counts nine advisory questions and Redbud counts ten, and the referendums overturning a rezoning are counted the same way by both. Wilmington's four sit on both lists. This is not a gap produced by differing criteria. It is six small-town rows that one list did not have yet. The issue is coverage, not definition.

The dashed box in the figure is the heart of this section. The Redbud tracker's methodology document sets down how its own registry grew. Its registry matches Ballotpedia's October 30 list of 41 measures in ten states, it writes, only after North Berwick, Peru, Wells, and Wiscasset in Maine, Harwood in North Dakota, and Downe Township in New Jersey were added on October 31. At least six of the rows the two lists share, then, are not a value two organizations reached by finding the same thing separately. One of them copied from the other.

So why was one list six rows short? The reason is neither criteria nor leanings. It is plumbing. One line the tracker document writes about Ohio sums up the situation.

Ohio's Secretary of State live results page posts statewide contests and State Issue 3, and nothing more. Gathering the state's twenty-one measures therefore means scraping fifteen separate county boards of elections, twelve of which publish results only as PDF summary reports. Three Kansas counties and two places in Illinois are no different. The tracker describes the state of this collection path as best effort with manual entry as the fallback. Maine poses another kind of difficulty. Every town posts its own warrant and counts its own ballots, so there is no central place to read from. That is why four of the six missing measures were small Maine towns.

Survey methodology calls the list you draw a sample from the sampling frame, and calls it undercoverage when that list omits part of the population outright. Undercoverage is awkward for two reasons. A unit absent from the list has zero probability of selection, which puts it in a different layer from a unit that refused to answer, and neither random selection nor a larger sample repairs it. Drawing at random from a skewed list produces a skewed sample. U.S. local governments numbered more than 90,000 units as of 2022, and even the Census Bureau puts work into keeping its list current as units are created and dissolved. Assembling a complete list on election day is a question of where to scrape from before it is a question of what to count.

When the list grew is worth recording too. The tracker document's table as of October 15 puts Ohio at twenty measures across seventeen jurisdictions; by election day it reads twenty-one across eighteen. It kept growing after that. Americus, Kansas, entered the registry on election day itself, November 3, and Piqua, Ohio, and Belvidere, New Jersey, arrived the next day, November 4. The denominator grew after the voting was over, which means every share published in the meantime was computed on a smaller one.

4.1Two Thirteens and Two Forty-Sevens

Both registries write "13 states," and the thirteens are not the same. Twelve overlap and one slot is off. Ballotpedia's thirteen includes Arkansas; Redbud's includes Georgia. The reasons differ as well. Clarksville, Arkansas, votes on its two measures on December 1, which places them outside a tracker whose range is cut at November 3 and earlier. Georgia's six are party-run advisory questions attached to the May primary ballots, which keeps them out of Ballotpedia's ballot measure classification. They are intra-party polls answered only by Republicans, or only by Democrats.

There are two 47s as well. Same figure, different bundle.

The split is over where the dates are cut. Ballotpedia counts by the year. It puts the measures reaching a ballot over data centers in 2026 at fifty for the year, with forty-one of them falling on November 3 alone. The other nine are six already held earlier in the year and three still ahead in December. That is how its scorecard as of November 4 reads forty-seven decided and three pending. The Redbud tracker cuts its range at the single day of November 3 and counts forty-seven inside it. One 47 is the number settled between the start of the year and election day; the other is the number that reached a ballot on election day.

Both columns say "47 measures, 13 states," and they do not point at the same bundle of events. Ballotpedia's six from earlier in the year are Canterbury, New Hampshire; Port Washington, Wisconsin; Monterey Park, California; Augusta, Michigan; Wrightstown, Wisconsin; and Nottingham, New Hampshire, and all six came out on the restricting side.

An article published two days after the election reduced this event to three numbers. ALEC posted it on November 5, 2026, under the title "After a Rebuke at the Ballot Box, What's Next for State Data Center Policy?", and the numbers are 13 states, at least 40 measures, and an average of 75%. We rebuilt all three from the raw files, from scratch. All three come out. One of them matches to the second decimal place.

Here are the three sentences as the article writes them. Voters in 13 states "considered at least 40 initiatives relating to data centers in some fashion." Then, "the position restricting data centers prevailed or was leading in 39 out of 40 measures." And finally, "On average, ballot measures restricting data centers prevailed with an overwhelming 75% of the vote."

We took the 59 rows of the Redbud registry and matched them one at a time. Below is that comparison. On the left is what the article says, in the middle is the set the value actually comes from, and on the right is our own computation.

Only the third line, "39 out of 40," sits on a different snapshot. The article recorded the measures that had prevailed or were leading as of November 5, 2026, while our figure splits the 39 rows with results into wins and losses in the copy received on November 7, Korea time. The other three reproduce as written. Source: computed directly from the Redbud tracker's registry and results files.

All three come out, and something still needs pointing at. What diverges is not the figures. It is that the set each sentence names and the set actually computed are two different sets. In two places.

First, the "40" is already a subset filtered to one direction. The registry runs to 59 rows, with 19 more allowing measures in it. The sentence reads "at least 40 initiatives relating to data centers," which sounds as though data center measures came to forty in all. Second, the "75%" is not the average of those 40. It is the average across all 54 rows with results, allowing measures included, and that works out to 75.01%. Averaging the set the sentence names, which is the 39 of the 40 restricting measures that have results in, gives 72.72%. The fortieth has no result yet and cannot enter an average.

Laid out together, every "average vote share" the same file can produce shows how far the choice of set moves the number.

Only the bottom line stands out high. Weighting by votes lets a single large-electorate row pull the whole figure up.

That bottom line is risky to lift on its own. The only Georgia row with results in is one advisory question from the DeKalb County Democratic primary, and it carries 146,998 votes, 34% of the 430,363 in the entire file. That row, at 94.4% on the restricting side, is what raises the weighted mean. With an intra-party advisory answered by Democrats alone holding a third of the "overall vote share," the figure cannot be carried over as the view of U.S. voters at large. The 72.90% restricted to November 3 is the one whose denominator can at least be named.

How the sourcing got folded is worth a line too. The article attributes its figures to an analysis by Data Centers on the Ballot built on Ballotpedia's measure tracking. The "13 states" and the "40," however, do not come out of Ballotpedia's list. Ballotpedia's thirteen has Arkansas and no Georgia, while that 13 has Georgia in it. It is the one slot from section 4. A three-step chain, Ballotpedia to Redbud to the citing article, was written down as a single step, and the denominator changed at the fold. This is compression of the chain, not false attribution. Redbud did draw on Ballotpedia, and that fact is itself what section 4 is about.

In Harwood, North Dakota, the one place where the data center side won, 430 people voted. 105 yes, 325 no. That is 0.17% of the 252,709 votes that came in on November 3. Without this figure alongside it, the headline's "1" reads as a larger share of forty-one than it is.

What was defeated there needs saying once as well. It was not a ban on data centers. It was a home rule charter amendment put up by residents gathering signatures, and its content is a procedural requirement. An industrial project that on its own exceeds peak demand of 100 megawatts, a 100-acre site, or 50,000 gallons of water a day would have to clear four things before approval. An independent expert review paid for by the applicant, disclosure of the record, a public hearing, and a written decision. Where several conditions overlap, the thresholds fall further. The definitions clause names data processing facilities explicitly.

The ballot summary itself never once says "data center." It records only that the amendment would place substantial additional requirements on the city for certain development and construction projects, and it directs voters to the full text printed in a local newspaper on August 25, 2026 for the particulars. It then records that the fiscal and operational impact of the amendment cannot be determined at this time. Voters were being sent to a newspaper from ten weeks earlier.

The share rejecting that measure is 75.6%, which is 325 divided by 430. The figure resembles the 75.01% from section 5 and points at the opposite thing. That 75.01% is the share the restricting side won on average nationally; this 75.6% is the share in Harwood that rejected a restricting measure. Set next to each other the two invert, so neither is written here without its subject attached.

Behind the rejection in Harwood sits an approval already granted. In 2025 the city approved Applied Digital's Polaris Forge 2 campus under a nondisclosure agreement, and residents organized afterward. The project was announced in August 2025 at roughly $3 billion and 280 megawatts. That figure moves with the moment too. 200 megawatts circulated on a load contract basis, 300 megawatts in an equipment supply announcement. The amendment does not reach an approval already settled; it applies to expansions and to projects still to come. A data center approved under a nondisclosure agreement becoming contested after the fact is the same structure this blog has set down twice, in the Pennsylvania case and in an audit of Amazon's disclosures.

Carrying this over as "Harwood residents voted for data centers" therefore gets it wrong. What was defeated there is a procedural requirement rather than a ban, and the ballot never used the words "data center." Going one step further, the sentence "the only place the data center side won" is itself a product of the direction label from section 3. The ballot box recorded the single word "defeated," and that defeat becomes a win for the data center side only if the row's direction label reads Restrict. Had the label gone on differently, the headline's "1" would have meant something else.

6.1Still Unsettled

As of November 7, 2026, Korea time, this event is still in motion. Four things remain open.

• Alhambra, California, is still counting. Mail ballots will be tallied for several more days. Los Angeles County's official certification date is November 27, 2026. The vote at this point is 83 to 17, and the 82 to 18 that circulated right after the election was an earlier partial-count snapshot.
• Boulder City, Nevada, has a settled outcome and vote totals that differ by source. The allowing measure was rejected and the restricting side took roughly 70%, and all three sources agree on that much. The total votes, though, come out as 7,160, 7,364, and 6,699, three different numbers. The result is usable; the vote count is not yet.
• Three more are scheduled for December. Clarksville, Arkansas, votes on two measures on December 1, and Mandan, North Dakota, on a ban above 25 megawatts on December 15. The two in Clarksville ask whether to keep an industrial development bond and a payment-in-lieu-of-taxes agreement, and the bond figure is a ceiling rather than a construction cost estimate.
• Passage is not the same as taking effect. The Emporia, Kansas, measure reached the ballot by court order and the city's validation suit is still live. Having passed, it takes effect only once there is a ruling. The reverse holds too. Had Alhambra's measure failed, the ordinance its council adopted in June would still be on the books.

In Ohio, the group that put the 25-megawatt threshold language on ballots in several jurisdictions is preparing a statewide constitutional amendment for 2027. It came up short of the signature requirement for the November 2026 ballot and is carrying the signatures it gathered forward to November 2, 2027. How that language was replicated is set down in the earlier article.

In preparing this article we counted the same data four separate times. Once by transcribing the full table and checking the arithmetic, once by running it from the raw files without copying any of those values, once while reconciling the two results, and once more before the body text was written. All four produced the same values. 41 rows and 252,709 votes, 27 passed against 14 defeated, 40 restricting against 1 development, 75.01%, 80.81%, and the 20, 44, and 46 that come of changing the reporting stage all agreed. The fourth pass also caught something, that two earlier documents had each recorded "72.72%, the average of the 40 restricting measures," when it is in fact the average of the 39 that have results.

One place did come out crooked, and the crooked side was our own computation path rather than the data. Ballotpedia's table prints each row's vote share rounded to a whole number, and averaging those forty-one integers as printed gives 73.56%. Recomputing the same forty-one rows from the vote counts gives 73.47%. The 73.48% in the section 5 table is that same average taken from the values the tracker stores to the decimal. The three sit within 0.09 percentage points of one another, so no sentence in this article changes. It goes down here for where it arose. This is a place where an already-rounded value was averaged a second time, and the story of this article, that deriving a value from a derived value drifts away from the original record, showed up at its smallest scale inside our own arithmetic.

Four counts agreeing four times is not a matter of our being careful. Countable data was public, and that is the whole of it. The Redbud tracker leaves its registry CSV, its results CSV, and its methodology document in the repository as they stand. On top of that it writes down, in figures, what it does not know.

Source: the Redbud tracker's methodology document and exclusion file (as received 2026-11-07 KST). The 39-of-47 verbatim rate and the five confirmed issue numbers were counted here directly, to check the document's description against the raw data.

That last row was not someone else's problem. When we read the tracker site by machine during the planning of this article, the screen we caught was exactly that preview. A figure reading 55% of simulated returns came back, and we knew it was not a real tally thanks to the label pasted across the screen in large type. No invented number made it into the article. The very accident the tracker worried about enough to label almost happened on our side, and the label is what stopped it.

What this tracker did already has a name. In machine learning there is a norm that a dataset should ship with documentation alongside it. The 2018 paper Datasheets for Datasets, by Gebru and co-authors, places these questions under a dataset's composition.

Complete or a sample; if a sample, what the larger set is; how representativeness was verified; if it is not representative, why not; and what was withheld. Five questions. The data statements proposal Bender and Friedman published the same year puts curation rationale first among its items, asking which material went in and what the goal was in choosing it. Both are norms written for machine learning datasets, with no election tracker in mind.

This tracker satisfies nearly all of those clauses regardless. It states that its list is not complete, the list having grown even the day after the election, says what the population is, marks which rows have no secured language, and publishes which cases it dropped, by name and with a reason. That is why opening it on the other side of the world three days after the election produces numbers that match, and why the places that do not match can be described down to where and why.

What this article did is the work DataClinic does on first entering a customer's data. Instead of taking the figures written in a report, it opens the original ledger, counts the rows, looks at the classification column, and hunts for rows that are missing. Doing that is how we found that 41 and 47 part ways over collection rather than definition, that 40 to 1 is a calculation rather than a record, and that the two lists are not independent of each other. A full row-by-row comparison is the kind of work nobody has to do by eye, which makes it a candidate for automation.

8.1Four Data Quality Lessons From This Case

Seen as a data quality problem, this case shows four things. Not one of the four is peculiar to election tallies.

• Completeness breaks before definitions do, and for duller reasons. The difference between the two lists was not a dispute over criteria but six missing small-town measures, and the cause of that omission was plumbing: with no central tally, fifteen separate county boards of elections have to be scraped one by one.
• One column of hand-written labels holds the whole conclusion. Flip the direction label and the headline flips with it. And that label went on from a single coder, which leaves its reliability unmeasurable.
• Recorded values and computed values in one table get mixed up by the reader. Pass and fail are what the ballot box wrote and direction is what a person added, and side by side in one table there is nothing to tell which of the two a figure was counted from.
• Two sources reporting the same number have not necessarily counted twice. Reading inheritance as cross-verification was the quietest trap in this event.

A dispute of the same shape is already running a second time inside this very subject, which is worth recording. Six organizations count how many U.S. local governments have moved on a data center moratorium, and their figures spread from 76 to 1,030. They part ways over whether to count attempts or only adoptions, whether to count temporary moratoriums alone or permanent bans as well, and whether one jurisdiction is one case or one action is. One of them splits its own figure across three stages: 1,030 places with a public meeting record, 632 that moved to a concrete action, and 270 that actually adopted. The three are not wrong against one another. They are different layers of one funnel, and quoting a single layer is the moment the layering disappears.

8.2Four Questions to Ask Before Shipping a Tally

Four questions come straight out of this article. They are not questions put to the number. They are questions put to the ledger that made it.

• Is this a recorded value or a computed one? If computed, does the computation rule ship alongside the metric?
• How was the list built, and what might be missing from it? Is there any way to gauge the size of what is missing?
• Who assigned the classification label, and would a second person assign the same one? If one person assigned it, does that fact sit next to the metric?
• Which moment is this figure a snapshot of? Does it take in rows still arriving, or only those already in?

The articles this blog has written along the same line mostly went after the method of measurement. Survey houses asked differently and the numbers parted ways, changing an item cut a trendline, changing the summary rule split the value, and the definition of revenue differed from company to company. What this article looks at is the step before that. The list of what is to be counted, and the label attached to each row. This one also closes in a different shape from the others. They closed on "so be careful with numbers," while here the ledger was open before any care was called for.

The figures in this article were downloaded from two public ledgers and computed here. Where the access path held a trap, the entry says so.

Primary — the two ledgers

• 1.Ballotpedia. "Data center-related ballot measures, 2026." Source of the 41-row November 3 table and the annual scorecard. As received 2026-11-07 (KST). ballotpedia.org
• 2.Redbud Consulting. "Data Centers on the Ballot." The page draws its data in the browser, so an automated fetch returns the simulated-returns screen. The real figures came from the repository files. datacentersontheballot.com
• 3.The same tracker's public repository. It holds the 59-row registry, the results file, the exclusion list of 10 measures it chose not to count, and the methodology document. Every recount in this article came out of these files. github.com/alexlundry/data_center_ballot_tracker
• 4.City of Harwood. Full text of the initiated home rule charter amendment (Article X). cityofharwood.com
• 5.North Dakota Secretary of State. Results page for the November 3, 2026 election. The Harwood measure is contest 1844. resultsnd.sos.nd.gov

Reporting and policy — where the headline numbers came from

• 6.CNBC (2026-11-04). Source of the 41 measures in ten states and of "all but one." cnbc.com
• 7.Scientific American (2026-11-04). Source of "41 times in 10 states" and of the split between the six earlier measures and the three in December. scientificamerican.com
• 8.American Legislative Exchange Council (2026-11-05). "After a Rebuke at the Ballot Box, What's Next for State Data Center Policy?" Source of the 13 states, at least 40 measures, and 75% average, and the home of the three numbers rebuilt in section 5. alec.org
• 9.The Washington Post (2026-11-04). Source of "at least 38 jurisdictions."
• 10.Stateline (2026-11-03, late local time). Source of the 33 passed or leading, and the earliest snapshot in this census.
• 11.The Epoch Times (2026-11-05). Source of the headline "38–1" and the body text's "at least 44." It quoted the tracker at an earlier snapshot.

Academic and normative — labels, sampling frames, documentation

• 12.Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, & Kate Crawford. "Datasheets for Datasets." arXiv:1803.09010 (Communications of the ACM 64(12), 2021). Source of the composition questions quoted in section 7. arxiv.org/abs/1803.09010
• 13.Emily M. Bender & Batya Friedman. "Data Statements for Natural Language Processing." TACL 6:587–604 (2018). The curation rationale item is paraphrased, not quoted. aclanthology.org
• 14.Klaus Krippendorff. Content Analysis: An Introduction to Its Methodology. Basis for the statement in section 3 that intercoder reliability coefficients require two or more observers.
• 15.Magnus Öberg & Mehmet Yilmaz. "Measurement issues in conflict event data." Research & Politics (2025). Source of the statement in section 3 that more than 90% of the fatality gap for Syria in 2017 traces to one auxiliary codebook rule. It sits behind a subscription, so the original could not be opened and it is paraphrased only. doi.org/10.1177/20531680251362440
• 16.ACLED. Comparing Conflict Data working paper (2019). Source of the multiple between ACLED and UCDP event counts.
• 17.The Washington Post. Fatal Force public data repository. Source of "more than twice the annual average" and of the FBI tracking about a third of fatal shootings as of 2021. Accessed 2026-11-07. github.com/washingtonpost/data-police-shootings
• 18.U.S. Census Bureau. Local government finance survey methodology and the 2022 Census of Governments. Source of the frame updating for units created and dissolved, and of the count of local government units.

Related Pebblous articles

• 19.Pebblous, "The Data Center Label OpenAI Wants to Drop Is Written Into the Poll Question" (2026-11-04). The preceding page of this article. report/openai-compute-campus-poll-wording
• 20.Pebblous, "Monterey Park Bans Data Centers by Ballot" (2026-07-16). blog/monterey-park-data-center-ban
• 21.Pebblous, "A Recount of METR's Experiment Finds No Ledger for Saved Time" (a case where the summary rule makes the number). report/metr-246-ai-time-saved-recount
• 22.Pebblous, "Did Data Center Opposition Grow, or Did the Question Change?" (2026-09-23). report/datacenter-opposition-reasons-pa-study
• 23.Pebblous, "Pennsylvania's Blanket Ban on Data Center Nondisclosure Agreements" (2026-08-21). blog/pennsylvania-data-center-nda-ban
• 24.Pebblous, "Amazon Drops Data Center NDAs, but Who Audits the Numbers?" (2026-10-05). blog/amazon-data-center-nda-disclosure-audit
• 25.Pebblous, "OpenAI and Anthropic Count Revenue Differently" (2026-10-10). blog/openai-anthropic-annualized-revenue-definition
• 26.Pebblous, "AI Hit a Wall This Summer, and It Wasn't Chips or Money" (2026-08-01). report/ai-datacenter-power-wall-2026