THE COUNTED CROWD

This summer, a research machine built by this studio read four hundred years of American history and came back with patterns.

Five hundred fifty-five dated instances, sorted into nine recurring shapes. Speculative manias, from Virginia tobacco to a vinyl doll. Body crazes, from the first fitness magazine to the weight-loss pill. And one pattern stronger than all the rest, everywhere the machine looked: the crowd against the institution. Farmers against banks, traders against hedge funds, debtors against courts. The machine found it in every century of American life, and when we sent it abroad, it found it in all six cultures it tested.

The pattern felt true. The feeling is the subject of this file.

Everyone knows the sensation. History rhymes, the saying goes, and the rhyme is a pleasure. Shays’ Rebellion in 1786 rhymes with the Whiskey Rebellion in 1794 rhymes with the farmers of the Grange in 1875 rhymes with the Populists in 1892 rhymes with Occupy in 2011 rhymes with a subreddit cornering a hedge fund in 2021. Read that chain once and you feel something click into place. Two hundred thirty-five years of the little crowd standing up to the big institution, wearing new clothes each time.

This studio has a standing rule about that click. The feeling of everything clicking into place is not the sensation of understanding. It is the sensation of a machine that has stopped listening.

So we did what the rule requires. We put the pattern on trial.

There is a serious record of American political instability. Peter Turchin, a quantitative historian, published a database of political violence in the United States from 1780 to 2010, some sixteen hundred episodes (his paper counts 1,590; the posted file has 1,602 usable rows): riots, lynchings, rampages, assassinations, insurrections, each one dated and sourced. He also published the shape of the record: a long instability wave across the second half of the nineteenth century, bracketed by two peaceful periods, with a fifty-year oscillation peaking around 1870, 1920, and 1970.

We tested the machine’s crowd pattern against that record twice. First against the instability peaks. Then against a second, gentler measure: ninety years of strike data.

The pattern lined up with neither test. Eleven of the chain’s sixteen instances fall inside Turchin’s years, and they scatter across the record the way random years would. Five fall inside the strike series, too few for that test to carry weight alone. The instances do not cluster in the violent decades or in the strike waves. The strongest pattern the machine found, the one that felt truest, tracks neither measure.

That does not show the pattern is false. It shows the pattern does not ride these two measures. A second finding sat in the chain itself, and it came from a question the machine was never asked.

Who got counted as the crowd?

Go back to the 1,602 rows of Turchin’s file and read the codings instead of the totals.

Forty-one percent of the events are coded to race. Forty-two percent of the deaths. The second-largest category of event in the entire record, after riots, is lynching: 438 episodes. The bloodiest twenty-year stretch of race-coded deaths in two hundred thirty years of data is 1860 to 1879, a window that takes in Reconstruction and is dominated by 1868. The file counts 1,576 dead there under its strictest “race” coding, 1,844 if combined codes such as “race, political” are included. Turchin leaves the Civil War itself out of the database.

And say plainly what that 438 is, because this file’s own argument requires it. The database counts episodes, discrete documented events in one research dataset built to measure instability, and it undercounts lynching deaths. The Equal Justice Initiative has documented 4,075 racial terror lynchings of African Americans in twelve Southern states between 1877 and 1950, a period its Reconstruction report puts at 4,500 lynchings; that report adds nearly 2,000 more between 1865 and 1876, for nearly 6,500 in all. Turchin makes the same point about his own file: for 1882 to 1964 he counts 451 lynching victims where Historical Statistics counts at least 4,745, and he calls the database’s lynching share “likely a serious underestimate.” He gives the reason: violence against Black Southerners was so routine that newspapers “simply did not bother to report most of these incidents.” The measuring stick has defaults too.

Now hold the record next to the machine’s beloved pattern, and watch the word crowd come apart.

In the pattern, the crowd is the hero. The farmers, the debtors, the traders, standing up to the bank, the court, the fund. But the largest single share of the actual American violence record is race-coded, and the database does not say who the crowd was: it has no field for perpetrator. Of its 649 race-coded events, 299 are lynchings and 325 are riots. EJI describes racial terror lynchings as “largely tolerated by state and federal officials.”

And the machine’s chain of heroic crowds, sixteen instances long, held zero Black-led protests. Not Nat Turner. Not the Exodusters. Not Montgomery. Not one. The chain skipped from the Dust Bowl to Occupy Wall Street: a seventy-six year hole with the sit-down strikes, the civil rights movement and the anti-war movement inside it.

This studio’s own test notes read the hole as a sign of what the chain sampled: “market- and consumer-adjacent” contention, “people with property enough to defend.” Why the sampling ran that way, this file has not shown. The likeliest reading is the archive’s own default about whose gatherings count as the crowd and whose get counted some other way: as a threat, as a disturbance, as a body count in somebody else’s ledger.

This archive has a name for an operation like that. In another file we watched a federal order rename the historical record, and the Park Service remove exhibit panels after. We called the mechanism what it is: the naming comes first, the taking comes second. If that reading holds, this is the same machine wearing laboratory clothes. Call it erasure by instance-selection. No one bans the history. It just never makes the dataset. And a pattern built on the dataset then circulates as four hundred years of truth, with the hole invisible, because a hole in a list of examples has no row to point at.

A blank is not neutral. A blank is non-coverage. So we stopped studying the hole and filled it.

Fifteen instances, with Birmingham and the March on Washington counted as one, drawn from institutional records and the press, among them state encyclopedias, university archives, the National Archives, the National Park Service, Stanford’s King Institute and The New York Times.

It starts at Stono, South Carolina, September 1739: as many as one hundred enslaved people, armed and moving toward Spanish Florida, in what the state’s own record calls its largest and bloodiest insurrection of enslaved people. Gabriel’s conspiracy in Richmond, 1800, betrayed before it moved; Virginia hanged twenty-six men to be sure. The German Coast of Louisiana, 1811, the largest revolt of enslaved people in American history, hundreds moving in formation on New Orleans. Charleston, 1822, thirty-five hanged on the word of a plan. Southampton County, 1831, Nat Turner’s revolt, named here plainly as what it was: an armed rising against the institution of slavery.

Hold the dates. The German Coast and Charleston fall inside what Turchin calls “a remarkably peaceful period in US history,” the first quarter of the nineteenth century, with Gabriel at its start. Turner’s rising came in what he calls the first spurt, in the 1830s. Stono predates his 1780 start. His calm is calm in recorded lethal violence, from a database he says “contains only a sample of all instability events.” Contention was a capital offense, and they organized anyway.

The chain runs on. Christiana, Pennsylvania, 1851: a free Black community mounts an armed defense against a federal posse enforcing the Fugitive Slave Act, and the treason prosecution that follows collapses. The Exodus of 1879, when Black Southerners answered Redemption not with a petition but with departure. The National Park Service puts the 1879 exodus at about six thousand. The census measures the whole decade: Kansas counted 17,108 Black residents in 1870 and 43,107 in 1880, and Kansapedia says “[l]arge numbers” came between 1879 and 1881. An exit, organized by the people exiting.

Lee County, Arkansas, September 1891: Black cotton pickers strike for a dollar per hundred pounds. At least fifteen of them are killed for it, strike leader Ben Patterson among them, several by extrajudicial hanging after capture. New York, July 1917: an estimated ten thousand people walk down Fifth Avenue in complete silence, the only sound the drums at the front of the line. Harlem, August 1920: the Universal Negro Improvement Association convenes delegates from twenty-two countries and holds a mass meeting at Madison Square Garden. The UCLA project editing the movement’s own papers says that “By 1920 the UNIA had hundreds of chapters worldwide.” The UNIA was a crowd insisting on being counted at scale, in a country that kept its count elsewhere.

Then 1941. A. Philip Randolph organizes a march on Washington and names the number out loud: ten, twenty, fifty thousand Black marchers on the White House lawn. President Roosevelt signs Executive Order 8802, banning discrimination in the defense industries, and the march never has to step off. Then Montgomery, 381 days of an emptied bus system. Greensboro, February 1960, four students at a lunch counter becoming seventy cities by April. Birmingham and the March on Washington in 1963, a quarter million people in front of the Lincoln Memorial. And 2020, when polls estimated that fifteen to twenty-six million people in this country said they had participated in the George Floyd protests. That is a survey of what people say they did, not a count of bodies in the street, and it should be said that way. The New York Times, reporting those polls in July 2020, said the figures would make the protests the largest movement in the country’s history. The Crowd Counting Consortium describes them as, at that time, the largest and broadest mobilization in U.S. history.

We ran the missing chain through the identical tests: same instability record, same strike series, nothing tuned. Thirteen of its instances fall inside Turchin’s years, and seven inside the strike series.

Against Turchin’s years it scatters too, a null like the machine’s chain. Against the strikes it leans the other way: five of the seven fall in the busiest quarter of strike years, too few to tell from chance.

The test cannot say why either chain scatters. What this chain does show is coverage: it appears in the violent decades and the quiet ones, under slavery and under Jim Crow and under the drone of the postwar boom. It was built to fill the hole, not sampled, so that coverage describes the record, not a law of it.

Turchin’s record calls the 1940s and 1950s the second peaceful period in American history. Richerson, summarizing Turchin’s Ages of Discord, describes the book’s account as reaching an age of concord from 1940 to 1960. Randolph’s 1941 ultimatum sits inside that calm. So do all 381 days of Montgomery.

So the pattern-hunters’ instinct, that crowds rise when history’s weather turns, does not fit these years. The 1941 ultimatum and Montgomery had no wave to ride.

· · ·

The machine that started this file was not malicious. That is what makes it worth a file. An archive-fed system, human or machine, can reproduce the archive’s defaults unless someone audits the rows. Ours may have. We found the hole because this house keeps a rule for it, and the rule is now yours.

The next time anyone shows you a pattern, in a book, in a feed, in a slide, from a machine, do not ask first whether the pattern is elegant. Ask who got counted. Ask what the missing rows would say. In 1917, an estimated ten thousand people walked down Fifth Avenue in silence. Read that the way this file reads it: a crowd making itself a row that could not be left out of the record.

Count again.

Correction

Corrected 2026-10-01. This essay published on August 6, 2026. The errors are named here, in fifteen entries, rather than quietly repaired.

Turchin’s record was described as long calm between three peaks. The essay said he published “peaks of instability around 1870, 1920, and 1970, and long stretches of calm between them.” His paper describes “an instability wave during the second half of the 19th century, bracketed by two peaceful periods (the first quarter of the 19th century and the middle decades of the 20th century, respectively),” with a fifty-year oscillation “with peaks around 1870, 1920, and 1970,” and says that “From 1860 to 1920, the level of violence fluctuated around a very high level.” The sentence now follows the paper.

Nat Turner’s rising was placed in the calm. The essay said the risings came “in decades when the national violence record was otherwise nearly quiet” and that “These risings happened in the calm.” Turchin calls the first quarter of the nineteenth century “a remarkably peaceful period in US history,” which holds for the German Coast and Charleston, with Gabriel at its start, but says “The first spurt occurred during the 1830s,” the decade of Turner’s rising, and his database begins after Stono. The passage now gives the dates.

Turchin’s count was given as 1,602. The essay said he “published a database of 1,602 documented episodes.” His paper says “The USPV database includes 1,590 unique instability events.” 1,602 is this studio’s count of usable rows in the posted file, excluding rows coded D or X. The sentence and the docket now give both.

The pull quote was attributed to Randolph alone. The pull quote was attributed to “A. Philip Randolph, 1941, the march that never needed to happen.” The National Archives quotes it as a threat by “Randolph, with others,” and Randolph’s wording as EBSCO Research Starters quotes him differs (“masses—ten, twenty, fifty thousand Negroes—on the White House lawn!”). The attribution now follows the Archives, and the docket carries the Archives’ sentence.

The Exodus figures were set against each other. The essay said the six-thousand figure “turns out to be a newspaper’s count of migrants passing through one city, St. Louis, in two spring months,” and that “The number that holds is the census.” No source this file read carries a 6,206 St. Louis count, and the National Park Service puts the 1879 exodus at “approximately 6,000.” The census measures the 1870s as a whole; Kansapedia says only that “Large numbers of blacks came between 1879 and 1881.” The passage now gives both figures and what each measures.

Every instance was said to be verified. The essay said “Fifteen instances. Every one verified against an institutional record: state encyclopedias, university archives, the National Archives, the Library of Congress. Every number below traces to a document a stranger can pull.” No Library of Congress record is among the sources listed for any instance; Birmingham had no source record, and the Garvey convention, the Exodus figure and the 2020 survey range rested on pages the essay did not list. The chain also names sixteen events if Birmingham and the March on Washington are counted apart. The sentence now names the record types and says how the fifteen are counted, and source records for Birmingham, the convention, the Exodus and the 2020 range are added.

The chain was said never to stop. The essay said “This chain scatters because it never stops.” Its own dates leave a gap from 1739 to 1800 and another from 1963 to 2020. The sentence is cut.

One reviewer’s summary was presented as the reviewers’ term. The essay said “The reviewers of his work name those decades an age of concord,” and the docket called Richerson’s text a review of the book with the dates as “1940-1960.” Richerson, summarizing Turchin’s Ages of Discord, writes “an age of concord (1940–1960).” The sentence and the docket now attribute it that way.

The lynching record was read as crowds doing the institution’s work. The essay said “in the largest single share of the actual American violence record, the crowd is not standing up to the institution. The crowd is doing the institution’s work off the books.” The database has no field for who the crowd was. Of its 649 race-coded events, 299 are lynchings and 325 are riots. “The lynch mob. The pogrom. The posse. Crowds enforcing what the courthouse wanted done but preferred not to sign.” is cut; the passage now gives those counts and quotes EJI on how officials treated the lynchings.

EJI’s counts were misadded and called a count of people. The essay said EJI had documented 4,075 lynchings “and nearly 2,000 more during Reconstruction. Nearly 6,500 documented killings,” and the docket said “A count of people, not episodes.” 4,075 and nearly 2,000 make about 6,075; EJI’s nearly 6,500 adds its Reconstruction count to the 4,500 its Reconstruction report gives for 1877 to 1950. EJI’s unit is lynchings, and no page this file read says it works “name by name.” The passage now gives EJI’s figures with the period each covers, and adds Turchin’s own comparison of his file against Historical Statistics.

Garvey’s biographers were cited without a source. The essay said “Garvey claimed two million members; his biographers do not believe him,” and “Even the exaggeration is evidence.” No source this file read supports either. Both are cut; the passage now quotes the UCLA project.

A decade and a half of discipline. The essay said the movement “had a decade and a half of trained discipline behind it” by 1960. None of its own dates gives fifteen years. The sentence is cut.

The 2020 superlative was stated without its “at that time.” The essay said “Even the floor of that range would make it the broadest mobilization ever measured in the United States.” The New York Times said the figures “would make the recent protests the largest movement in the country’s history,” and the Crowd Counting Consortium says “at that time.” The sentence now follows both.

The 1860 to 1879 window was called Reconstruction and mixed two codings. The essay said the window was “the years the country calls Reconstruction” and counted “roughly 1,576 dead in race-coded violence alone.” Reconstruction ran from 1865 to 1877. The 1,576 covers only events coded “race,” while the essay’s 41 and 42 percent count every event whose code includes race, under which the window holds 1,844. The sentence now gives both and says 1868 dominates the window.

The argument claimed more than its test shows. The essay read the same null result two ways: of the machine’s chain, “A pattern that tracks nothing is not a discovery,” and of its own, “Here the null is the finding: continuity, not cycle.” It called the machine’s seventy-six-year hole an inherited archive default (“That is the finding”), and said the movement’s organizers “had been keeping the same discipline since 1739. It always was,” that “Organizing in the calm is what made the storm survivable,” that “The count alone moved the institution” in 1941, and that the chain was “the deepest lineage of protest in the country” missed by “the archive’s oldest habit.” It also called the machine’s pattern one that “tracks nothing at all,” felt true “while being so empty,” and said “Nobody programmed that hole. … The machine was not told to leave anyone out,” that “The foundation of the modern movement was poured in the stillest water of the American century,” and of the machine, “Ours did.” The essay never gave the sizes of its tests: eleven of the machine’s instances against the violence record and five against the strike series, a test this studio’s own notes call underpowered. Of the missing chain it said “It scatters too,” when against the strike series five of that chain’s seven instances fall in the busiest quarter of strike years, a lean too small to tell from chance. This studio’s own test notes read the hole as a sign that the machine’s chain sampled market- and consumer-adjacent contention, and the hole also skips the sit-down strikes and the anti-war movement. EBSCO’s account of 1941 says Roosevelt feared “serious violence” and answered with “a mix of concessions and arm-twisting.” Each of these lines is cut or narrowed; the essay now gives the test sizes, says the test cannot say why either chain scatters, and offers the hole’s cause as the likeliest reading, not a finding. The studio’s notes are filed as a source record; they are not public.

Smaller changes made at the same time:

• “Peter Turchin, the field’s leading quantitative historian” now reads “a quantitative historian.”
• The excerpt said the pattern “failed every test, and the reason is a lesson in who gets counted”; the test result and the missing rows are two findings, and the excerpt now says so. “it found it on every continent it checked” now reads “in all six cultures it tested.”
• “And inside the continuity, one fact worth the whole file” is cut.
• “with the test design registered in writing before any result was computed” is cut. The strike test’s design is time-stamped before its run in this studio’s repository; the first test’s design is attested only in its notes.
• The paragraph on the edge of the calm, the 1960 sit-ins and the 1963 campaigns, is cut. The conclusion it set up is among the lines cut above.
• A sentence now says Turchin leaves the Civil War out of his database, as his paper states.
• “the closest thing the record has to a national ledger of ordinary people organizing” is cut from the description of the strike data.
• “It is not a count of the dead” now reads “it undercounts lynching deaths”; the file records fatalities. Turchin’s own reason for the undercount is now quoted.
• “Every dataset encodes its builder’s questions, including the one this essay leans on” is cut.
• “In another file we watched a federal order rename the historical record before removing pieces of it” now reads “rename the historical record, and the Park Service remove exhibit panels after.” That file’s own correction says the order names no panel, photograph or trail, and that the Park Service removed the President’s House panels.
• “Then 1941, and the cleanest demonstration of crowd-against-institution in the century” now reads “Then 1941.”
• “ten thousand people” in the silent parade now reads “an estimated ten thousand,” as the Yale record gives it.
• “fills Madison Square Garden” now reads “holds a mass meeting at Madison Square Garden.”
• “Every archive-fed system, human or machine, will reproduce the archive’s defaults” now reads “An archive-fed system … can reproduce.”
• The source records for Turchin’s paper, Richerson’s article, EJI, the Kansapedia entry, the UCLA project and the National Archives repeated several of these errors; each is corrected.

Sources: Peter Turchin, “Dynamics of political instability in the United States, 1780–2010,” Journal of Peace Research 49(4), 2012, and the USPVD2010 dataset, both read through the Internet Archive (the author’s links are dead). Peter J. Richerson, Cliodynamics 8(2), 2017. Equal Justice Initiative, Lynching in America and Reconstruction in America. National Archives, Executive Order 8802. EBSCO Research Starters. Kansas Historical Society, Kansapedia. National Park Service, Homestead National Historical Park. UCLA Marcus Garvey Papers Project. PBS American Experience. The New York Times, July 3, 2020, read through the Internet Archive. Harvard Ash Center, Crowd Counting Consortium. Stanford’s King Institute. This studio’s own test notes and run logs for the pattern machine and the missing chain, which are not public. Each read on October 1, 2026.

Corrected 2026-10-03. This essay published on August 6, 2026. One sentence still carried a claim the 2026-10-01 correction had withdrawn. It is named here rather than quietly repaired.

A registration claim the earlier correction withdrew. The essay said: “We ran the missing chain through the identical tests, registered in writing before they ran: same instability record, same strike series, nothing tuned.” The 2026-10-01 correction found that the strike test’s design is time-stamped before its run in this studio’s repository, and that the first test’s design is attested only in its notes. “Registered in writing before they ran” claimed for both tests what the record supports for one. It now reads: “We ran the missing chain through the identical tests: same instability record, same strike series, nothing tuned.”