Who Kept the Books

The mechanism

Whoever holds money has to describe it to someone: a crown, a legislature, a court, investors, the public. The holder speaks first because the holder has the papers and a reason to speak. Anyone with a different account has to get the papers before saying so, and that takes access, time and a forum the holder does not run.

Nobody has to lie for this to work. A planter filing a low count, a secretary of state calling a gold shipment protection, and a company calling a chat-deletion default ordinary practice may each believe what they wrote. The record still leans toward the account that serves its keeper, because the keeper decides what gets written down, what gets kept and when it is shown. The Haitian historian Michel-Rolph Trouillot named four points where that happens: when sources are made, when archives are assembled, when facts are retrieved into a story, and when a later age decides what mattered (documented fact that he argues this; high; Silencing the Past, chapter 1). The cases below sit at each of those points. Sections 1 to 7 follow the money from Hispaniola to Iran-Contra, section 8 tests the same question on AI firms and their customers, section 9 records two popular claims that failed the test, and section 10 states what a record cannot settle.

Does not prove: that any holder in this piece acted in bad faith. The mechanism needs no intent, and the cases were not chosen to test intent.

1. The count made by the party it pays

In 1514 the Spanish crown counted the Taíno people of Hispaniola in order to assign them to Spanish holders. The demographic historian Massimo Livi-Bacci calls it "a population count that served as the instrument of population relocation" (documented fact that he writes this; high; Livi-Bacci 2006).

The count covered only people already held in encomienda. People in flight stood outside it. Who counted as "Indian" was a social label: the same child of a Spanish father and a Taína mother might be listed as Spanish if raised Catholic and Castilian-speaking in a town, and as Indian if raised in a village (contested account, moderate; Lynne Guitar, Kacike, 2002).

Later sugar-estate counts on Hispaniola show the same habit. In 1530 they list 200-plus Indians and 700-plus unspecified "others"; in 1533, 200-plus Indians and 1,525-plus "others" (documented in the returns Guitar read; moderate). Planters had written to the crown swearing that their encomienda people were dead, in letters written to obtain licenses to import enslaved Africans. The archbishop wrote that the uncounted people were mostly Indians. Guitar reads the repeated round "200" as a number nobody wanted to state exactly (contested account, moderate).

Estimates of Hispaniola's population at contact run from about 100,000 (the historian Rosenblat) to 7 or 8 million (Cook and Borah), with Las Casas's polemical 3 million between them. Livi-Bacci's estimate is a few hundred thousand, falling below 20,000 by late 1518 (high that he states both; the other figures are contested or polemical, moderate that their authors proposed them). The surviving counts were made by the system's operators, and they left out the people who fled and the people relabeled.

The people counted did not end with the count. Taíno descendants live today in Cuba, Puerto Rico and Hispaniola. The Smithsonian's National Museum of the American Indian documents surviving traditions and present Taíno identification (documented fact; high). José Barreiro, a Smithsonian scholar emeritus and Taíno descendant, describes a kin network of more than 4,000 people in Cuba's eastern mountains (contested account, moderate; another source gives about 20,000, and the two figures conflict). When Haiti declared independence in 1804, it took the island's Taíno name, Hayti (documented fact; moderate).

Bearing (inference, moderate). The record shows who gained from a low number and who kept the books. It does not show what any one planter knew. The same question applies when a lab runs and reports its own evaluation: who made the count, who gains from its result, and what sits in the bin labeled "other". It applies to outside evaluators too, and to the verification tools the author of this series builds.

Does not prove: intent to deceive in any single return, that any lab or evaluator misreports results, or that a model evaluation resembles colonial coercion. The parallel concerns who keeps the record. It carries none of the conquest's harm.

2. The remedy that moves the cost

Bartolomé de Las Casas came to Hispaniola as a settler and held Indigenous people in encomienda before he became a Dominican friar and the conquest's best-known critic (documented fact; high). In his Historia de las Indias (book 3, chapter 129) he gives his own account of 1516 to 1518. Colonists who wanted sugar mills asked for licenses to bring enslaved Africans, and some promised him they would free the Indigenous people they held if they could import about a dozen each. He obtained the king's consent. The Council set a first license for 4,000 people across four islands. More licenses followed: by his estimate over 30,000 enslaved Africans to Hispaniola and over 100,000 to the Indies. A water mill needed at least 80 workers and a horse mill 30 or 40, and he notes that the king's revenue grew with the mills. The Indigenous workers never received the promised freedom. He later judged himself guilty, by inadvertence, once he saw that African captivity was as unjust as Indigenous captivity (documented fact that he wrote this; high).

Mechanism (inference, moderate). The fix kept the revenue model and changed who carried the cost. Las Casas names that chain himself. The fullest account of it in this record is the one Las Casas, who brokered the remedy, wrote later in his own history. The Atlantic trade already ran through Portuguese traders and African sellers, whom he also describes. Naming the sellers puts no blame on Africans as a people; the demand in this chapter came from the colonists' mills and the crown's revenue. Nor does his account license any judgment of Spaniards or Catholics as a people. His Brief Account became the base of the anti-Spanish "Black Legend", and the conquest's sharpest critics in this record were Spanish Catholic friars.

The AI half is a question, with no AI record behind it. When a lab announces a safety fix, who now absorbs its cost: reviewers, contracted labelers, users in a language the evaluation never covered? This research found no record showing that any present fix moved its cost this way. The question stays a checking question until someone finds one.

Does not prove: that Las Casas caused the Atlantic slave trade, or that any present safety fix has displaced harm.

3. An exact ledger for an unjust claim

Haiti won its independence by revolution in 1804. France recognized it on a price. A royal ordinance dated 17 April 1825 set 150 million francs, payable in five annual installments, with a customs advantage for French trade. Haiti's Senate accepted on 11 July 1825, after a squadron of 14 French ships arrived (documented fact; high for the terms, from the Haitian historian Gusti-Klara Gaillard-Pourchet in an essay published by the Bibliothèque nationale de France; high for the ordinance date and moderate for the acceptance date and the squadron, from reference and press accounts).

The accounts are exact. To make the first payment Haiti borrowed 30 million francs in France at 6 percent over 25 years and received about 24 million after deductions. In 1838 the total fell to 90 million over 30 years, and the former colonists' share was cut by 40 percent. The money went to about 8,000 former colonists or their heirs and to the creditors who held the loans. The valuation used enslaved people as its unit of account, at 2,000 to 4,500 francs per person (documented fact; high; Gaillard-Pourchet).

Different obligations ended on different dates, and a single "final payment" year hides that. The indemnity itself was settled in 1878 through refinancing; the 1825 and 1875 loans were repaid in 1887 to 1888 and 1921 to 1922 (documented fact; high that Gaillard-Pourchet gives these dates). Another source gives 1883 for the last indemnity payment, and this piece leaves that conflict open. Some accounts carry the debt to 1947, when bonds from a 1922 refunding loan were redeemed. That reading treats a later US-arranged refinancing as a continuation of the first burden. It is defensible, and it is still an interpretation (low to moderate).

Gaillard-Pourchet's arithmetic documents a transfer of resources. It confers no moral right to demand them. An account can be exact and the obligation it records can still be unjust, and nothing in the accuracy of the ledger settles that question.

On 17 April 2025, the two-hundredth anniversary of the ordinance, President Emmanuel Macron said the indemnity had put a price on the freedom of a young nation, and he announced a joint commission of French and Haitian historians. He announced no reparations (official claim; high). This research found no published commission report by 1 October 2026 (unknown, which is different from none).

Does not prove: that the indemnity alone explains Haiti's later poverty. That claim needs a counterfactual economic model, and none is offered here. Domestic political responsibility and later foreign intervention stay in the account.

4. Custody without consent

On 17 December 1914 the US gunboat Machias carried $500,000 in gold from Haiti's national bank in Port-au-Prince to New York. Secretary of State William Jennings Bryan's letter of 31 December 1914 to Haiti's minister gives the official account: the bank's officers "made application" for the gunboat, the move protected American interests, and the gold was "deposited there at interest". The same letter records Haiti's protest that its sovereignty had been violated (documented fact; high; Foreign Relations of the United States 1914, document 583). The letter names no receiving bank. Secondary accounts disagree about whose vault it reached, and this piece leaves the bank unnamed until those accounts are read in full.

Marines landed the next July. Through the occupation of 1915 to 1934 the United States controlled Haiti's finances by treaty, and a 1918 constitution allowing foreign land ownership passed by plebiscite after the Assembly was dissolved (documented fact; high; Congressional Research Service report 94-459). The evidence table below carries the rest.

Mechanism (inference, moderate). The people holding the asset and the people it belonged to had come apart, and the holders decided. The transfer record is accurate. A reader who takes only the official justification loses the owner's objection, which sits in the same letter. Later, treaty terms and bond covenants kept financial control in place after the troops left.

Does not prove: that the transfer caused the occupation, which has documented strategic motives beside commercial ones; that any one banker caused it; or that data held by an AI lab compares with a seized national reserve. The 1914 transfer was a military act against a sovereign state.

5. An official account, and the part that waited forty years

The account the Katangese government gave out from 10 to 13 February 1961 said Patrice Lumumba and two fellow prisoners had escaped, been caught by villagers and died violently. He was executed on 17 January 1961. A United Nations commission of inquiry concluded in 1961 that the escape story had been fabricated. In 2001 a Belgian parliamentary commission of inquiry, set up after a 1999 book by Ludo De Witte raised the question of Belgium's responsibility, reported that the official version was a lie (documented fact; high; the commission's report, volume I, read directly). It found that Belgium bore moral responsibility for his killing (documented fact; moderate, read through a summary, because the conclusions volume could not be fetched). The first account did not stand: outsiders doubted it at the time. Belgium's own part in the killing waited forty years for its parliament. The record on Lumumba also includes the army campaign he ordered in South Kasai, which was accompanied by massacres of Baluba civilians (documented fact; moderate).

The United States Senate's Church Committee had earlier examined the American side. Its interim report on assassination plots found that lethal substances reached the CIA's Congo station, found no evidence that they were used, and drew no connection between them and his death (documented fact, a congressional finding; high). The same committee's staff report on Chile from 1963 to 1973 found US covert involvement "extensive and continuous". It printed Richard Helms's notes of 15 September 1970, which list $10,000,000 and the instruction to "make the economy scream". It traced corporate money too. After John McCone called Helms, the telecommunications firm ITT's chief Harold Geneen offered money to the head of the CIA's Western Hemisphere division. ITT later offered $1 million, which was refused, and after the CIA advised it on channels ITT passed at least $350,000 to the Alessandri presidential campaign (documented fact, a congressional finding; high).

Mechanism (inference, moderate). In Chile the money moved through agency and company books that the public could not see, and the record moved when a Senate committee with document access obtained those books. In the Lumumba case a United Nations inquiry rejected the first account within the year, and the record of Belgium's part moved only after an author's book led Belgium's parliament to open its own inquiry, four decades after the event.

Does not prove: any equivalence between these acts and anything a present-day company does, or that the CIA caused Lumumba's death. The committee's own finding says it found no evidence the substances were used.

6. The inquiry needs access and time

On 13 November 1986 President Reagan told the nation, "We did not ... trade weapons or anything else for hostages." On 4 March 1987 he said his heart still told him he had not traded arms for hostages, but the facts and the evidence said otherwise, and he called it a mistake (official claims; high; presidential addresses). The independent counsel, Lawrence Walsh, later reported that more than $47 million moved through the private "Enterprise" accounts, and that support to the contras from October 1984 to October 1986 violated the Boland Amendment ban. He found that documents were "systematically and willfully withheld from investigators" (documented fact; high; Walsh, Final Report, executive summary).

Where access or time ran out, the record stopped short. CIA director William Casey was hospitalized "with a fatal illness before Independent Counsel was appointed", and Walsh wrote that "no formal investigation of Casey was ever undertaken". Pardons on 24 December 1992, including one for Caspar Weinberger, ended prosecutions (documented fact; high). Walsh could not prove that Reagan knew of the diversion of funds to the contras, and this piece does not claim he did. A review given little time shows the same limit. Attorney General Barr appointed the retired judge Frederick Lacey on 16 October 1992 to review the Justice Department's handling of the BNL case, over guaranteed credit to Iraq through one bank branch (see the evidence table), with a report due on 8 December, about six weeks. Lacey found no independent counsel was needed. Representative Henry Gonzalez, who chaired House Banking, called the review a "whitewash", said key documents on one charge were never requested and noted that an appendix stayed classified (the dates and the conclusion as his floor statement gives them, read directly, high that he states them; his charge is a contested account; Congressional Record, 21 January 1993).

Outside inquiries also clear people. In 2002 Vice President Cheney said there was "no doubt" Iraq had weapons of mass destruction, and President Bush said Iraq was "reconstituting" its nuclear program (official claims; high). The Iraq Survey Group later found the programs had ended between 1991 and 1996 (documented fact; high; Duelfer report, key findings). In 2004 the Senate Intelligence Committee found that most key judgments in the intelligence estimate were overstated or not supported by the underlying intelligence (Conclusion 1). In the same report it found no evidence that administration officials tried to "coerce, influence or pressure analysts" (Conclusion 83) (documented fact; high for the text as NBC News published it in full). In Britain the Chilcot inquiry reported in 2016 that the threat judgments had been presented "with a certainty that was not justified", that the flawed assessments behind policy were not challenged when they should have been, and that planning for after the invasion was wholly inadequate. It found that the Joint Intelligence Committee should have made clear the intelligence had not established beyond doubt that Iraq had continued to produce chemical and biological weapons (documented fact that the inquiry's public statement of 6 July 2016 says this; high, read directly).

Bearing (inference, moderate). An outside review with real access and time can convict or acquit, and these records show it doing both. A review whose scope the reviewed party sets can do neither with authority. Who Knew First records the one AI incident review with access to an operator's logs published by 23 September 2026, METR's review of the OpenAI and Hugging Face activity. OpenAI set its window at 26 June to 13 July, which left out an alignment flag of 25 May. It requested one of the seven questions, kept redaction rights, first supplied about 285 selected transcripts and gave feedback that led to edits, and the reviewers spent about $400,000 of OpenAI credits. Anthropic's METR review has terms known only from Anthropic's account, and METR's published terms for that review would be the same test applied to the compiler's maker (documented fact as Who Knew First reports it; high). On 30 September 2026, in written testimony to a Senate subcommittee, METR's Chris Painter said the review was "scoped to a limited set of questions" (official claim by METR; high, read directly).

Does not prove: that any window was chosen to leave out a warning, or that any conclusion changed; the record shows neither, and METR stands by its conclusions. Nor does it prove that officials knew their claims were false (Conclusion 83 points the other way on pressure), that any AI evaluator lacks integrity, or that any present review arrangement resembles Iran-Contra conduct. The point concerns access and time.

7. The backup that broke the deletion

Between 22 and 29 November 1986, Oliver North deleted 736 messages from the White House's PROF electronic mail system, and John Poindexter deleted 5,012. The White House Communications Agency kept routine backup tapes and had retained the set from 15 November. Investigators recovered the messages from them. Poindexter also destroyed the only signed presidential Finding meant to authorize the November 1985 HAWK missile shipment after the fact (documented fact; high; Walsh, Final Report, chapter 3).

The deleters controlled their own accounts. They did not control a copy that an office nobody was thinking about had made by routine, before anyone knew it would matter.

Bearing (inference, moderate). This is the property Who Knew First asks of an outside incident ledger: a copy held by someone other than the party the record could embarrass, made on a schedule that party does not control. A cryptographic hash can show that a record has not changed since it was hashed, and only against a trusted earlier hash. It says nothing about a record that was never written. Only a copy held outside can show that a record was not removed.

State archives show what happens without that copy, at each of Trouillot's points. British colonial offices kept restricted "watch" files, and the 1961 instructions known as Operation Legacy told governors to separate files for destruction or removal and to leave no trace of the category (documented fact, high, from the historian Tim Livsey; the instruction's wording moderate). In 2011, in a suit by Kenyan survivors of the 1950s detention camps, the Foreign Office acknowledged undeclared files held at Hanslope Park, later counted at about 20,000 files from 41 territories. In 2013 the government said it sincerely regretted the abuses and agreed a settlement worth £19.9 million, costs included, for 5,228 claimants (documented fact, high; the payment figure from the Foreign Secretary's statement to Parliament of 6 June 2013, read directly). South Africa's Truth and Reconciliation Commission found that destruction of state records was a coordinated effort sanctioned by the Cabinet, systematic from 1990 to 1994 and most intense in 1993, and judged it probably the single factor that most undermined its investigations. The same chapter says the National Intelligence Service headquarters alone destroyed about 44 tons of paper and microfilm records in six to eight months of 1993 (documented fact, high; TRC final report, volume 1, chapter 8, read directly). The CIA destroyed 92 videotapes of two detainees' interrogations in November 2005, and a special prosecutor closed his inquiry in 2010 without charges. The Senate Intelligence Committee's full study of CIA detention remains classified; only a 525-page summary is public (documented fact, high). One report says the returned copies sit with the Senate outside the Freedom of Information Act (moderate).

Two limits travel with these cases. In Kenya the files were not the first route to the truth: survivors' memoirs and testimony had documented the abuses, the historian David Anderson built the statement that forced the 2011 acknowledgment from public Foreign Office files of 1967, and Caroline Elkins judged that the released files added volume without changing what was known (documented fact, high, as Livsey reports it). In South Africa, what was destroyed is gone, and no later inquiry can close that part of the gap (inference, moderate).

Courts can also stand in for the outside copy. In Epic Games v. Google the judge found chat deletion and told the jury it could infer the deleted messages would have been unfavorable to Google, and the jury found for Epic on 11 December 2023 (documented fact, high; instruction wording moderate). The tobacco industry's internal documents opened through litigation: Minnesota's 1998 settlement required public depositories, and a library is reported to hold over 19 million of them (documented fact for the route, high; the count moderate).

Does not prove: that backups always serve accountability. The same property cuts the other way: backups also hold data that people have a right to have deleted, and an order to preserve everything can override those deletions. On 13 May 2025 a court ordered OpenAI to preserve output logs that would otherwise be deleted, including chats users had deleted, and the going-forward duty ended on 26 September 2025 (moderate, from secondary summaries). Intent can also fail as proof: the Supreme Court unanimously reversed Arthur Andersen's conviction for shredding Enron documents under its retention policy, because the jury had not been required to find consciousness of wrongdoing (documented fact, high; Arthur Andersen LLP v. United States, 2005).

8. The AI-era record

The modern cases are smaller and different in kind. Each carries part of the same mechanism, and each comes with the point where the comparison breaks.

Three AI firms with a court record on their own data. In United States v. Google, Judge Amit Mehta found in his opinion of 5 August 2024 that Google had, since 2008, set internal chats to delete after 24 hours unless history was turned on, and kept that default after litigation holds. He wrote that the court was "taken aback by the lengths to which Google goes" to avoid creating a paper trail. He declined sanctions because the deleted chats would not have changed the liability finding (documented fact; high; as the ABA Journal quotes the opinion). A competing explanation stays beside it: short retention defaults also cut storage costs and privacy risk, and the judge did not need to resolve intent. In the consolidated copyright litigation against OpenAI, Magistrate Judge Ona Wang's order of 24 November 2025 records as undisputed that an OpenAI employee downloaded books from the LibGen shadow library in 2018 and that OpenAI deleted two book datasets in 2022, about a year before the suits began. OpenAI first gave non-use as the reason and later said the reasons were privileged. The order found that OpenAI had waived that privilege and rejected the plaintiffs' crime-fraud theory (documented fact; high; the order's text). In Bartz v. Anthropic, Judge William Alsup ruled on 23 June 2025 that using the books to train Anthropic's models was fair use, as was scanning print books it had bought, and denied Anthropic summary judgment on the more than seven million pirated copies it had downloaded for a central library, setting that question for trial. The parties settled for $1.5 billion, Anthropic denying any wrongdoing, and the court granted final approval on 20 July 2026 (documented fact; high; the order and the court-approved notice, read directly). In each case a court record, and not the firm's own account, set out how the records or the data were handled.

A fourth record reached the public through the press. In 2024 OpenAI's departure paperwork tied lifelong nondisparagement terms to former employees' vested equity. After press reports, OpenAI told former employees on 23 May 2024 that it would not cancel vested units and would release the obligations unless they were mutual (documented fact; high).

Does not prove: that Google's default hid any specific conduct; spoliation or bad intent by OpenAI, since the court found waiver and made no spoliation finding; that any suppressed account of OpenAI existed; or any deletion or concealment by Anthropic, which no source read alleges.

Routing past a legal bar. Since October 2022 the US has required licenses to export advanced AI chips to China. Three Justice Department cases since then follow the Iran-Contra pattern of intermediaries and false papers. In one, announced on 8 December 2025, a defendant and his company pleaded guilty to smuggling and unlawful export; the department says they exported or tried to export at least $160 million of Nvidia H100 and H200 chips, falsified shipping papers and received more than $50 million in wire transfers from China. In a second, four people are charged with using a Tampa company with "Realtor" in its name, which never did real estate, as a front to move 400 A100 chips to China through Malaysia and Thailand. In a third, an indictment of 19 March 2026 alleges that about $510 million of a US server maker's equipment went to China in a few weeks of 2025, and that the defendants staged thousands of non-working "dummy" servers to pass the maker's own compliance audit and a Commerce Department inspection (documented fact for what the department states; high. The plea is a conviction; the rest are allegations until proved).

Where it breaks: in Iran-Contra the officials who made policy ran the scheme. Here the charged people are private brokers and employees, and the department describes the server maker's compliance team as a target of the deception. Does not prove: that any chip designer, cloud firm or lab knew of the diversions, or that the charged people are guilty. Responsibility in these cases attaches to the people charged, never to Chinese people or any other group.

Oversight behind spending, and a negative result. The Commission on Wartime Contracting found that at least $31 billion, and possibly as much as $60 billion, of more than $206 billion in contracts and grants in Iraq and Afghanistan was lost to waste and fraud (documented fact; high). On the AI side, GAO report GAO-26-107859 of 13 April 2026 found that the Defense Department, Homeland Security, the General Services Administration and Veterans Affairs were not systematically collecting lessons learned from their AI acquisitions. All four agreed with its recommendations, and its sample of 13 acquisitions was chosen to be nongeneralizable (documented fact; high). The July 2025 defense notices read for this piece, Google's and Anthropic's among them, each pair a $200 million ceiling with $1,999,998 obligated at award, about 1 percent of the headline figure (documented fact; high; the percentage is arithmetic). This research found no AI-side record of measured waste or fraud, and GAO reports no waste figure. The oversight gap is documented and a measured loss is not. Inspector general reports on AI contracts were not searched, so a wider search could change this.

Does not prove: that any AI contract wasted money, that the agreements are good or bad value, or that the ceilings will be reached.

A provider that learned from the press. In September 2025 Microsoft cut an Israeli military unit, Unit 8200, off from specific Azure cloud services after press reports on how the unit used them. The company's president, Brad Smith, said Microsoft would not have known to look into the matter without that reporting (documented fact that he said it; high; TechCrunch, 25 September 2025). By the company's own account, the outside reporting is what prompted it to look.

Where it breaks: the review was the company's own, of its own customer, and no public body ruled. This piece names one unit and one company. Other providers' and other states' defense cloud contracts, Project Nimbus among them, were not searched, so the choice carries no judgment that this is the worst case. Does not prove: any use of the services in targeting, or anything about data Microsoft says it did not access.

9. Two popular claims the record does not support

"The CIA created or paid bin Laden." The 9/11 Commission found that bin Laden and his comrades received "little or no assistance from the United States" (documented fact; high; 9/11 Commission Report, chapter 2). The documented risk is narrower: Pakistan's intelligence service chose the recipients of the program it distributed.

"The Clintons stole billions meant for Haiti." FactCheck.org found on 18 October 2016 that the emails cited gave no evidence of recovery contracts steered to Clinton friends or donors, and that the State Department had found no evidence of preferential treatment. The Washington Post Fact Checker and PolitiFact rated related claims false or mostly false (investigated and not supported; moderate). The documented criticism of 2010 aid stands on its own records, in the table below.

These negative results stay in the piece. A record that names every gap it can document should also name the gaps it looked for and did not find.

10. Two problems, two answers

Haiti's ledger shows what records cannot do. The indemnity accounts were exact and the claim they enforced was unjust. France's acknowledgment in 2025 was a political act, and no record system could have produced it.

The same discipline runs the other way for anyone who argues from money. Follow it far enough to find the contract, the decision and the people who bore the loss, then check the result. A funding tie justifies scrutiny and proves no bought conclusion. A missed target can establish failure without establishing fraud. GAO's 2015 review of 23 key US reconstruction activities in Haiti after the 2010 earthquake, each above $10 million, found reduced planned outcomes in five of six infrastructure activities and three of seventeen others. It also found that the non-infrastructure activities met or exceeded half of their performance targets, and mission officials told GAO that thin staffing and unrealistic first plans caused delays (documented fact; high; GAO-15-517). An argument that drops either half of that record is less true.

The engineering answer this series argues for covers the part a record can reach: copies held outside the party under review, counts an outsider can redo, disclosure clocks set outside the holder, and a named owner for each review. As a set they are proposals. This piece did not check which labs use any of them, and the series does not claim they work at scale. Existing clocks carry a limit of their own. California's SB 53 sends critical safety incident reports to the state Office of Emergency Services within 15 days, exempts them from the Public Records Act, and gives the public only anonymized aggregates from 2027, though members of the public may also file reports. The EU AI Act routes serious-incident reports for high-risk systems to market surveillance authorities. A clock routed to a closed recipient moves the custody question one step: the regulator then chooses the summaries the public story is told from (documented fact for SB 53, high; for the EU text, moderate; the custody point is inference, moderate). The industry's own published record is thinnest on training and use: the 2025 Foundation Model Transparency Index found training data and post-deployment use among the least disclosed areas, with an average score of 40.69 and only 7 of 23 contacted companies submitting reports (documented fact, high; its authors warn that method changes limit comparison across years). What they cannot do is choose a just rule. Who profits, who pays and who can be made to answer are set by contract and law, and changing them is work for institutions, courts and legislatures. The criterion the history supports is narrower than either: the people who bear the consequences of an account should be able to inspect its basis, challenge it in a forum the holder does not run, and obtain a remedy when it fails (inference, moderate).

How this connects to Who Knew First

Who Knew First infers that the party holding an AI incident's logs also names it, and that the name sets how fast anyone else hears about it. This piece shows the older form of that pattern at the scale of states: whoever controlled the money wrote the first account, and the record moved only when an inquiry, a court, an archive or a reporter got hold of it. It also explains that op-ed's proposed remedy from a different angle. A 1986 deletion failed because a routine backup sat with an office the deleters did not control, which is the property the op-ed asks of an outside incident ledger. The scale and the harm here are of another order. The pairing transfers a question about records, never the weight of the wrong. It also transfers no verdict: these historical cases do not show that an AI operator never widens its own first account, and whether one does is a question for the incident record itself. The Terms for Telling, a companion piece, follows the people inside who could tell, including the 1972 Tuskegee disclosure and the 2005 release of edited climate reports.

Evidence table: what was said, what the record shows

Eighteen cases in rough date order, each setting a public statement or stated purpose beside the documented trail. The last column is the limit that travels with each row.

What the table shows (inference, moderate). In each documented case, the party that controlled the money also controlled the first public description of it, and the gap closed only after someone outside that party obtained the record. None of these cases shows a controlling party closing its own gap before an outside party forced the record open. Two archive cases in section 7 qualify this: in Kenya, survivors' testimony documented the abuses before the files opened, and in South Africa destruction left part of the gap open for good.

What this does not prove

• No intent. Nothing here shows that any present AI company, lab, evaluator or regulator acts as the states, officials or firms in these records acted. The parallels name mechanisms, and each states where it breaks.
• No equivalence. Software evaluation is not enslavement, occupation or war. Each parallel carries a question about records and authority, never the moral weight of the historical harm.
• No frequency. The cases were chosen because their gaps are documented. They cannot show how often holders disclose first, how often outside forums win, or that no holder ever closed its own gap.
• No proof that engineering suffices or fails. The instruments in section 10 are proposals, and this piece did not check which labs use any of them. Whether they improve decisions at a tolerable cost is an open, testable question.
• No settled history. Several load-bearing items are moderate or contested: Guitar's reading of the planters' motive, the 1914 vault, the dates of Haiti's 1825 acceptance, the 2010 aid shares, the Lumumba inquiry's wording, the Saudi file. The unread sources listed below could move them.
• No representative selection. The intelligence cases center on the United States and the United Kingdom because those states ran public inquiries. That is an effect of which records exist. It is no finding about other states, and a missing record stays unknown, never evidence of innocence or guilt.

Open threads

These are leads for anyone who wants to dig. Each says what is known, what is missing and what it could change.

1. Stated limits loosened before defense awards. OpenAI's usage policy dropped its "military and warfare" category in January 2024; Google's AI Principles dropped their "applications we will not pursue" section in early 2025; Anthropic announced contractual exceptions for selected government agencies in June 2024. All three later received defense agreements with $200 million ceilings (OpenAI's read only through summaries). The policy texts and Anthropic's announcement were read directly. OpenAI's and Google's stated reasons were read only through search summaries. Needed: The Intercept and CNBC reports of January 2024 and Google's February 2025 post, opened directly. Could change: whether the three moves read as one pattern, and how each was explained. No record read shows that any policy changed because of a pending contract.
2. Distribution through platforms the vendor does not run. Defense models reach users through integrators and agency platforms, and Anthropic reported in November 2025 that it had detected an espionage campaign, which it attributes to a state sponsor, using its coding tool. Outside security researchers questioned the report. Needed: the researchers' criticism read directly, any Anthropic reply, and a reader outside Anthropic's model family for the Anthropic row. Could change: whether a vendor's account of its own product's misuse can be checked by anyone else.
3. Control of the checking forum across ideologies. Kwame Nkrumah, Thomas Sankara, Fidel Castro and Iran in 1953 each offer a record on who controlled the forum that could check the state. Needed: each leader's record carried whole, including the repression the fairness review added (Nkrumah's dismissal of a chief justice after a treason acquittal; Sankara's mass dismissal of striking teachers and executions after a hasty trial; Castro's forced-labor camps and the jailing of dissidents in 2003). Could change: whether the mechanism holds across political systems or only in the cases read.
4. Who carries the cost. Britain's Anatomy Act of 1832 gave anatomists unclaimed bodies from workhouses, hospitals and prisons; data labelers at the contractor Sama, paid as little as $1.32 an hour after tax (TIME, January 2023), offer a modern case. Needed: the Act's terms from a primary text (the current source is an encyclopedia) and research on data-consent law, which has not been done. Could change: whether section 2's question has an AI-side record behind it.
5. Readings this piece rests on. A Haitian or Dominican scholar's account beside Guitar's on the Trujillo-era revival of indio color terms; a same-method case drawn from another state's declassified files. Could change: whether the intelligence cases apply one standard in practice.
6. Direct reads still owed. Each could raise or lower a label above:
   • The UN Special Envoy for Haiti's own aid tables (the 2010 aid shares).
   • Peter James Hudson's chapter on the 1914 gold (the receiving vault and bank).
   • The membership and any output of the Franco-Haitian historians' commission.
   • The New York Times "Ransom" series methodology note (its present-value estimates).
   • The Tower Commission report (when the backup tapes were found and used).
   • Noble David Cook's 1993 work on disease in the collapse of Hispaniola's population.
   • The 2020 Census counts of people identifying as Taíno, read at census.gov (two unverified figures conflict).
   • Caribbean-authored archaeology and history (Frank Moya Pons, Roberto Cassá, Jalil Sued-Badillo).
   • Trouillot's Silencing the Past, chapters 2 and 3 in full.

How this was made

An Anthropic-built model, Claude Opus 5.5, compiled the research and drafted this piece at the author's request. The research ran in passes on 28 September and 1 October 2026: four source studies, a fact check of 266 claims that found 12 wrong or unsupported and corrected all of them, a fairness check that corrected 13 framing problems, a second check of the AI-side pairings, and an independent check of the first pass in a separate review session. Every check was run by AI models at the author's request. Anthropic appears in the record, so a check by a model from the same maker is a known limit. None of this is an outside review. A reader who finds an error can send it, and it will be corrected in the dated section below. A pass on 1 October 2026 added passages carried from the same checked research files (the archive and court cases in section 7, the Chilcot and Lacey reviews, the METR terms as Who Knew First publishes them); it did no new research. Before publication on 8 October 2026, these primary sources were read directly: the Bartz order and the settlement notice, the Belgian commission's report (volume I), the Chilcot statement, the TRC report's chapter on records, the 2013 statement to Parliament on the Kenya settlement, Gonzalez's floor statement and METR's testimony. They raised several labels, corrected the description of the Bartz ruling and rewrote the Lumumba section, which had said the first account stood for forty years. The Belgian conclusions volume refused the fetch, and the other unread sources stay listed in the open threads.

Corrections

None yet. Each correction will be listed here with its date and what changed.

Sources

Primary records and official inquiries

• Foreign Relations of the United States, 1914, document 583 (Bryan letter, 31 December 1914): https://history.state.gov/historicaldocuments/frus1914/d583
• Office of the Historian, Haiti country record: https://history.state.gov/countries/haiti
• Congressional Research Service, report 94-459, Haiti (26 May 1994)
• Élysée, declaration of the President on France and Haiti, 17 April 2025: https://www.elysee.fr/emmanuel-macron/2025/04/17/declaration-du-president-de-la-republique-sur-la-relation-entre-la-france-et-haiti
• Lawrence Walsh, Final Report of the Independent Counsel for Iran/Contra Matters, executive summary: https://irp.fas.org/offdocs/walsh/execsum.htm, and chapter 3 (same mirror)
• Ronald Reagan, addresses of 13 November 1986 and 4 March 1987 (Reagan Library; American Presidency Project)
• Senate Select Committee on Intelligence, S. Rept. 108-301 (2004), conclusions, full text as published by NBC News: https://www.nbcnews.com/id/wbna5403731; Phase II release, 5 June 2008
• Iraq Survey Group (Duelfer) report, key findings (National Security Archive copy)
• Commission on Wartime Contracting in Iraq and Afghanistan, final report (2011)
• 9/11 Commission Report, chapter 2: https://9-11commission.gov/report/911Report_Ch2.htm
• Senate Select Committee to Study Governmental Operations with Respect to Intelligence Activities (Church Committee), interim report "Alleged Assassination Plots Involving Foreign Leaders" and staff report "Covert Action in Chile 1963-1973" (intelligence.senate.gov)
• National Security Archive, Electronic Briefing Book 82 (US policy in the Iran-Iraq war); Riegle report, 9 February 1994
• GAO-15-517, Haiti reconstruction: https://www.gao.gov/products/gao-15-517
• GAO-26-107859, Artificial Intelligence Acquisitions, 13 April 2026: https://files.gao.gov/reports/GAO-26-107859/index.html
• Department of Defense contract announcements, 14 July 2025: https://www.war.gov/News/Contracts/Contract/Article/4243673/
• US Department of Justice releases of 20 November 2025, 8 December 2025 and 19 March 2026: https://www.justice.gov/opa/pr/us-citizens-and-chinese-nationals-arrested-exporting-artificial-intelligence-technology ; https://www.justice.gov/opa/pr/us-authorities-shut-down-major-china-linked-ai-tech-smuggling-network ; https://www.justice.gov/opa/pr/three-charged-conspiring-unlawfully-divert-cutting-edge-us-artificial-intelligence
• OpenAI copyright multidistrict litigation, S.D.N.Y. No. 25-md-3143, Opinion and Order of 24 November 2025 (Magistrate Judge Ona T. Wang): https://cases.justia.com/federal/district-courts/new-york/nysdce/1:2023cv08292/606655/782/0.pdf
• Bartz et al. v. Anthropic PBC, N.D. Cal. No. 3:24-cv-5417, Order on Fair Use, 23 June 2025 (docket 231): https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0.pdf ; official settlement site and court-approved notice: https://www.anthropiccopyrightsettlement.com/
• Truth and Reconciliation Commission of South Africa, Final Report, volume 1, chapter 8: https://www.justice.gov.za/trc/report/finalreport/Volume%201.pdf
• The Iraq Inquiry (Chilcot), Sir John Chilcot's public statement, 6 July 2016 (the inquiry's site, archived copy): https://web.archive.org/web/20171205050213/http://www.iraqinquiry.org.uk/the-inquiry/sir-john-chilcots-public-statement/
• Rep. Henry Gonzalez, "Update on BNL Investigation", Congressional Record, 21 January 1993 (FAS mirror)
• Arthur Andersen LLP v. United States, 544 U.S. 696 (2005)
• Foreign Secretary William Hague, statement to Parliament on the settlement of Mau Mau claims, 6 June 2013: https://www.gov.uk/government/news/statement-to-parliament-on-settlement-of-mau-mau-claims
• California SB 53 (2025), Business and Professions Code 22757.13, leginfo.legislature.ca.gov
• Senate Select Committee on Intelligence, CIA detention and interrogation study, executive summary, 9 December 2014

Scholarship

• Massimo Livi-Bacci, Population and Development Review 32(2), 2006
• Lynne Guitar, Kacike, December 2002
• Bartolomé de Las Casas, Historia de las Indias, book 3, chapter 129 (Duffy translation)
• Gusti-Klara Gaillard-Pourchet, Bibliothèque nationale de France, France-Amériques: https://heritage.bnf.fr/france-ameriques/en/haitis-independence-debt-slave-unit-account-1794-1922
• Tim Livsey, History Workshop Journal 93 (2022), on colonial "watch" files and the migrated archives
• Wan, Klyman, Kapoor and colleagues, 2025 Foundation Model Transparency Index, arXiv 2512.10169
• Michel-Rolph Trouillot, Silencing the Past, chapter 1 (NYU Press excerpt): https://nyu.manifoldapp.org/read/silencing-the-past-excerpt-trulliot/section/2f251038-7193-4e95-aba1-01953ed871c3
• José Barreiro, American Indian magazine 18(4), Winter 2017
• National Museum of the American Indian, Taíno gallery guide: https://americanindian.si.edu/nk360/pdf/Taino-Gallery-Guide-English.pdf

Journalism, fact-checks and reference

• Haitian Times and Associated Press, 17 April 2025 (France's acknowledgment)
• NPR, 16 June 2016 (Senate staff report on the Red Cross); ProPublica and NPR, 3 June 2015
• PolitiFact, 11 October 2016 (Caracol)
• Just Security, 1 December 2016, and UN News, June 2017 (cholera)
• FactCheck.org, 18 October 2016 (Haiti aid claims)
• ABA Journal, August 2024 (the Mehta opinion in United States v. Google)
• NBC Los Angeles and CNBC, 23 May 2024 (OpenAI's exit terms)
• TechCrunch, Rebecca Szkutak, 25 September 2025 (Microsoft and Unit 8200)
• Who Knew First, on METR's review terms (harperz9.github.io)
• METR, Chris Painter's written Senate testimony, 30 September 2026: https://metr.org/blog/2026-09-30-chris-painter-senate-testimony/
• Legal Dive, December 2023 (Epic Games v. Google instruction and verdict); CNN and NPR, November 2010 (CIA tapes inquiry)
• Summaries of the 13 May 2025 and 9 October 2025 preservation orders in the OpenAI copyright litigation (Huntress, Legal500)
• Wikipedia, "FCO Migrated Archives" and "Truth Tobacco Industry Documents", used for counts at moderate confidence
• ProPublica, 10 September 2026, and Transnational Litigation Blog, 11 September 2025 (the 9/11 Saudi file)
• Wikipedia, "Assassination of Patrice Lumumba" and "Bartz v. Anthropic", used for dates and as an index to the sources they cite, never as sole support for a high label
• Belgian Chamber of Representatives, parliamentary inquiry into the assassination of Patrice Lumumba, report, DOC 50 0312/006, volume I, 16 November 2001: https://www.lachambre.be/FLWB/PDF/50/0312/50K0312006.pdf