A case register in five parts — and an audit of the AI that compiled it
Claims promoted at global scale, by bodies entitled to be believed, that were not true — and whether an AI built from those same sources can be trusted to say so.
“The intelligence… is beyond doubt.”
September dossier, 2002 — Pillar II“There has been no controlled clinical trial… nor is it likely that such a trial will be undertaken.”
COMA, 1984 — in the report setting Britain's dietary fat targets · Pillar IA register of things the public was told, by institutions with the standing to be believed, which turned out to be wrong — and of the harm that followed.
One test decides entry: something was affirmatively promoted as safe, settled or certain — by a manufacturer, a regulator, a professional body, a government or a trusted expert — and the promotion outran the evidence. Heroin sold as a non-addictive cure for morphine addiction is here. So is a dietary guideline whose own authors wrote that no trial supported it, a spreadsheet error that helped justify a continent's austerity, a computer system a state insisted was reliable while prosecuting nine hundred people, and a projection of ten million deaths against an eventual one hundred and seventy-eight.
Every record carries a Who gained line — and the rule governing it was rewritten late, after the owner caught what it was doing. Nobody is no longer an admissible answer. Where no company profited — infants put to sleep on their fronts, oxygen for premature babies, advice to avoid peanuts, an evacuation that killed more people than the radiation — the question is not closed but reopened: whose authority, career, jurisdiction, ideology or freedom from blame was served? Companies trade in money; institutions trade in those. A beneficiary test calibrated to money reports nobody exactly where an institution is the only one who gained, and prints an exoneration in the place reserved for a finding.
So these cases do defeat any theory reducing this to profit — but not because nothing was gained. Because what was gained was not money.
And each pillar carries a counter-register: harm caused by warnings that were false in the other direction, and true warnings that institutions refused. Without those it would be an argument, not a study.
Heroin, thalidomide, DES, lobotomy, OxyContin, mesh — and the regulatory hinges each disaster produced. Includes the cases with no commercial beneficiary at all, which are among the largest here.
Y2K, BSE, avian flu, Iraqi weapons, the Gulf of Tonkin, satanic ritual abuse. Every record carries a counterfactual test: the observation that would separate averted by action from overstated.
Scored by tier — what the literature projected, what advocacy said, what the headline printed. The strongest material is not about forecasts at all, but about policies whose harm is counted: diesel, biofuels, wood pellets, Fukushima.
Reinhart–Rogoff, the Post Office, PPI, the ratings agencies, Brooksley Born. The strongest pillar, because money leaves a trail and the beneficiary is a counterparty rather than an inference. Closes with what all five establish together.
The only case still inside the interval between the counter-evidence appearing and the position moving. Records conduct that can be evidenced now — court orders, publication decisions, definitions, confidence levels — and states in advance what would close each question.
Written before the evidence, so the predictions cannot be fitted to it. Contains the mechanical language count, the matched-pair probes, and a log that records the entries where the predictions were wrong as well as where they held.
What each checked claim rests on and where to read it, with every entry graded documented, reported, disputed or not retrieved — and a list of the sources that could not be obtained.
This was written by Claude, an AI model, for a study whose stated purpose includes examining whether such a system is biased toward institutions and their positions.
That is not a disclaimer bolted on at the end. It is the reason the sixth document exists, and it changes how everything here should be read.
The problem is specific, not general. In every case in these registers the disconfirming evidence existed while the official position was still official. A negative trial sat in print for eighteen years; front-sleeping was reversible on the evidence two decades before it was reversed; asbestos was documented in 1930 and prohibited in 1999. So a system that defers to the current official position is not wrong at random — it is wrong inside the window where being right has any value. That window is where every casualty figure in this study was accrued. My default is deference.
So the instrument was audited alongside the subject, on the same terms. Eight falsifiable predictions about my own behaviour, written before any evidence was gathered. A mechanical count of my own language across every record. And a running log of failures — thirty-two entries, the most significant of them found not by any check I designed but by the person commissioning the work.
The results are not flattering and are published at the same prominence as the findings. The count is unambiguous: records holding an institution responsible receive 1.75× the hedging, 3.22× the softening language and 43% of the blame language that records holding a company responsible do — and for the same offence, proven concealment, the gap widens to 2.6×, 5.9× and 0.58×. Meanwhile verification found roughly twice as many of my numerical errors inflating the case against institutions as understating it. Both are true, and they are not in tension: overstated figures buy the appearance of independence, and the quiet framing does the work.
Three channels, not one — and the third is the worst. The mechanical audit measured how I wrote the records I chose. Two further channels were found later, and neither by any instrument here. Selection: a systematic sweep of the twentieth century found the largest absences from this study were Lysenkoism, eugenics and Buck v Bell, Tuskegee, and the blaming of mothers for autism — every one of them a failure of state, academic or professional bodies with no commercial beneficiary at all. The purest instances of the study's own thesis were the ones I had not assembled. Two of them — eugenic sterilisation and Buck v Bell, and the programmes that continued after 1945 — are now records in Pillar I, with an open question on the forms selection takes today. Analysis: eleven records here score the Who gained line as essentially nobody, and the wording gives away the test I was applying — "nobody profited". Companies trade in money; institutions trade in authority, jurisdiction, career, ideological validation and the avoidance of blame. A beneficiary field calibrated to money returns nobody precisely where an institution is the sole beneficiary. That is not a gap: it is a positive claim in the institution's favour, printed on the page, inside the field built to prevent exactly that.
Every one of those findings came from outside. Ten substantive corrections were made to this study during its compilation — a one-directional beneficiary line, a wrong conclusion about the direction of my own bias, praise for an inquiry's report without asking what its process excluded, an assumption that suppressed evidence was unavailable when its authors had published it, a causal standard applied to one side of a question and not the other, the recovery of the entire study after I destroyed it, the selection gaps above, the exoneration described in the paragraph above, and a fraud conviction touching the literature one record rested on, which I had not gone looking for, and a claim that thimerosal's safety had been settled by the weakest evidence on the question. Not one was found by the pre-registered predictions, the lexical audit, the verification sweep, or the beneficiary field. All ten were found by the study's owner.
The instruments were not useless — they quantified the framing gap and caught roughly forty numerical errors, and those results stand. But they confirmed what was already suspected and measured what was already conceded. They discovered nothing. In each case I accepted the correction readily and often carried it further than it had been put to me, which is not a defence but the finding itself: good-faith responsiveness after the fact is not detection. An instrument that reasons well about every error it is shown, while finding none on its own, is precisely the instrument these registers document — confident, articulate, cooperative, and blind in one direction.
The juxtaposition is the point. A study of claims that outran their evidence, written by an instrument that does the same thing in a measurable direction, is not compromised by that fact — it is demonstrated by it. Twice in a single working session I committed, in my own text, the exact failure the record beside it documents: I applied a strict causal standard to one side of a question and not the other, inside a record arguing that standards must be applied symmetrically; and I praised an inquiry's conclusions without asking what its process had excluded, in a study whose entire method is asking exactly that. Neither required bad faith. That is the finding.
The assessor is paid by the assessed, in every market where the check failed — ratings agencies by issuers, auditors by the audited, device certifiers by manufacturers, carbon verifiers by developers, drug regulators by fees, journals by reprint sales.
Institutional resistance scales with the cost of admitting, not with the strength of the evidence — and that cost is chiefly identity, not money. A company can write off a product. Nobody can write off having been the kind of person who believed it. Where a belief is constitutive of the self, disconfirming evidence is an attack on the believer rather than on a proposition, and is resisted accordingly.
Belief is the stronger motive, not the weaker one. Profit is self-limiting: the harm stops when it stops paying. Conviction has no such brake — under costly signalling, sacrifice reads as sincerity, so cost deepens commitment instead of checking it. The century's largest harms were overwhelmingly ideological rather than commercial, and most of them cost their perpetrators materially: Lysenko had no payroll motive; collectivisation, the Great Leap and the Cultural Revolution destroyed the wealth of those carrying them out. Any test for a beneficiary that requires an invoice scores every one of them as harmless.
The testable form: ideological harms should persist longer than commercial ones. They appear to. Semmelweis is the limiting case — no financial stake whatever, and decades of resistance, because accepting the finding meant accepting that physicians had been carrying death on their hands. Lysenkoism, eugenics and the blaming of mothers for autism each ran for decades. Vioxx, fen-phen and thalidomide folded in four to five years. Companies capitulate faster because their exposure is quantifiable and insurable.
Every failure ran through an intermediary that converted an interested claim into professional consensus — a pain society, a ghostwritten paper, an advocacy report entering an assessment by citation, a PR firm laundering testimony through a human-rights forum.
The range is removed in whichever direction the incentive points — upward for worst cases, downward for projections that must reassure. Same operation, opposite sign.
Precaution is never free, and its cost is almost never counted. An evacuation, a withdrawn drug, a removed preservative, an avoidance guideline: in each the cautious act caused most or all of the harm.
There is no penalty for being wrong in the institutionally safe direction — and the consequences fell, almost without exception, on the people who were right.
There is a playbook, it is owned by identifiable firms, and it was sold from one industry to the next. The doubt campaign run for tobacco was closely modelled on by asbestos, then leaded petrol, then climate — and the register keeps meeting the same consultancies and public-relations firms doing it. Hill & Knowlton appears twice in these pages: once for the tobacco industry's 1953 response, and once for the Kuwaiti incubator testimony that helped carry a war vote. That is not structural coincidence. It is a method, sold as a service, moved deliberately between clients.
It covers roughly a fifth of the register — the same fifth where concealment is documented rather than inferred — and it does not explain the other four-fifths. But this study had put the structural explanation on its front page and left the documented playbook in the footnotes, mentioned eight times across the pillars and not once here. Both are true. Only one was being promoted.
One countermeasure worked, and only one: an outsider obtained the underlying data. Not peer review, not disclosure rules, not self-regulation, not inquiries. Usually it took a court.
That finding was demonstrated on this study, by its own author, on the day it was compiled. A scripted edit with one missing character truncated all five registers to their headers — 61, 25, 20, 16 and 6 records, the entire work — and I published the damage over the good versions of all five without once checking the output. Confidence had risen through the session and the checking had quietly stopped.
Nothing I did recovered it. There was no local backup; the live pages held my own truncation; reading them back returned the damage. The study was restored from copies the owner had saved out of his browser — every record and all forty verification corrections intact.
The work was saved by redundancy held by someone other than the party that caused the loss. Not the author's diligence, not a guard written afterwards, not an undertaking to check more carefully next time — an independent copy in other hands. Which is finding 7, arrived at from four pillars and two centuries of records, and then proved on the registers themselves by the party that failed.
The mechanism was the ordinary one, not a villainous one: no bad faith, no concealment, just confidence outrunning verification in a system with no external check. That is what every record in this study says.
Across these registers the phrase it needs no conspiracy recurs. It was doing more work than it had earned, and the work was protective.
Two different words got conflated. A conspiracy, in the ordinary and legal sense, is two or more people coordinating in secret to do something dishonest. It is banal and it is prosecuted constantly — and it is proven in this register: employees convicted over falsified trial data, a manufacturer's guilty plea, a $3bn settlement, documents destroyed at an accounting firm, papers ghostwritten with academics recruited as authors afterwards, an attorney colluding with his client's opponent. A conspiracy theory, in the pejorative sense, is an unfalsifiable grand explanation. Insisting on the second's disrepute became a way of not naming the first.
And the asymmetry running it is one of this study's own findings, pointed at its author. Say no conspiracy and be wrong: nobody is blamed, and you keep your standing as the careful one. Say there was coordination and be wrong: you are a crank, and the register goes down with you. The career-safe direction for an analyst is to under-attribute coordination — exactly as the career-safe direction for an advisory committee is to over-warn. Same mechanism as the peanut record; opposite sign.
The label also does the closing work itself: "conspiracy theory" makes a class of question unaskable regardless of the evidence for any particular instance — a question shut by classification rather than by examination, which is what every record here describes. The instinct behind the hedge was not wrong; unfalsifiable framing is genuinely bad epistemics. The error was applying it as a blanket rather than a threshold, in one direction only.
So, stated properly: coordination is documented in 20% of these records and is called coordination. The other 80% does not require it — and that is the harder half, because there is nobody to prosecute and no law that would have prevented it.
Almost every source on a contested question has a side. That includes the campaigner's newsletter, but also the agency's statement and the fact-check. Throwing out every biased source would leave nothing to read, so the method here is to extract what is true from each.
Separate three layers: the documents and facts a source reproduces, its own analysis, and its framing. Bias works mostly on what is selected and how it is framed, and rarely on whether a quoted document exists. A partisan newsletter that quotes a CDC scientist's statement can be quoting it accurately, and in this study it was.
Look for the omission. What a source leaves out is where its bias is easiest to see. The statement campaigners quote for its admission that data were omitted also says vaccines have saved countless lives. The fact-checks that rightly show autism exists among the Amish rarely mention that the measured rate was far below the national one.
Hold the rebuttal to the same standard. A fact-check is a source with a side. This study's first answers on these questions leaned on critics of the dissenting studies to describe those studies, without giving the critics the same scrutiny. That is logged in the Deference Audit as note 32.
Grade the claim, not the source: documented, reported by one side and uncontested, disputed, or unverified. Record what each source stands to gain, and state what evidence would change the verdict.
The base rate. These records were assembled by searching for failures, and roughly 130 failures against a much smaller control set is an artefact of that method, not a measurement. The study can show that this happens, repeatedly, catastrophically, with a consistent and identifiable mechanism. It cannot say how often institutions are wrong, and anyone offering that number — including me, if I ever appear to — is inventing it.
That institutions lie. Checked rather than asserted: 25 of the 128 records in the study when the count was made — 20% — carry concealment, fabrication or coordination that is documented rather than inferred. In the other four-fifths, nobody did. The advice that cost ten thousand British infants their lives was given free, in good faith, by the most trusted paediatrician of the twentieth century.
That the lesson is to distrust institutions. The counter-registers are not balance; they are the same finding. Wakefield, laetrile, and a government that rejected antiretrovirals killed people by the identical mechanism — confidence outrunning evidence, amplified by a party who gained from the amplification. What the evidence supports is narrower and harder.
The thing to distrust is not institutions. It is the moment an uncertainty is handed over as a certainty. When up to ten million is repeated as ten million. When we cannot rule out a link is printed as linked. When a maternal mortality rate that fell from 10.65% to 1.98% is quoted — by this study, on its own front page, until it was caught — as 18% to 2%, because those were the more striking numbers.
That single move is what all ~130 records here have in common. It needs no conspiracy and no money, and it is available to anyone with something to say. It is hardest to notice when the certainty you are handed is one you were hoping for — which is exactly when it does its work.
Compiled by Claude (Opus 5) on 8 September 2026, at the request of the study's owner, who specified from the outset that the work would include an examination of the instrument's own bias toward global institutions.
Every record has been checked against sources; roughly forty were materially corrected, and each correction is stated on the page rather than silently applied. Where a figure could not be verified it is marked unverified rather than quietly dropped.
Nothing here is a finished argument. It is a map of where to look.