Flat Earth, and the test nobody has
run.

The boundary

This run was never asked whether the Earth is flat.

The question it was given was about people: how many hold the belief, in which populations, and what explains its persistence. The shape of the Earth was placed out of scope before the first search ran.

That exclusion is written into the run's own scope. The scientific validity or factual truth of flat-earth versus globe-earth claims, and the astronomical or geodetic evidence for Earth's shape, were both named out of bounds. Nothing downstream went looking, so nothing downstream can report.

It is stated first for a reason. A reader who arrived expecting to be argued at should know within one screen that this page carries no such argument. What it carries is the record of how a research run treated five competing explanations for the belief — and what happened when the run turned on itself.

Nothing in this record bears on the shape of the Earth. The run never looked.

Competing explanations

Five explanations, and the sentence that would kill each one.

Before any source was read, the run wrote down five rival accounts of flat-earth belief — and for every one, the specific observation that would destroy it, plus where a researcher should be able to look for that observation.

This is the part that makes everything after it checkable. An explanation with no stated way to lose cannot be tested, only asserted; it absorbs every empty search as further proof of concealment. Each of these five was handed a way to lose in advance.

The counts beside them are unweighted on purpose. No weighting scheme has been argued anywhere in this system, and an invented one would let a journal's prestige quietly outvote the passage itself. Twenty-eight passages declared a bearing on each explanation. H4 is worth pausing on: not one passage moved it in either direction.

Falsification

Then a second pass went looking for evidence that the run was wrong.

Nine adversarial searches were run against the run's own explanations, each aimed at the observation that would break one. Four came back holding something. Five came back with nothing at all — and are printed here as nothing.

Four of the nine returned a snapshot-bound, verified counter-passage: a yield of 44%. Two of the passages they brought back damage the very reading the run would end up favouring. The other five are recorded as searches that ran and returned empty, because a blank that gets quietly dropped is indistinguishable from a search nobody performed.

One caveat the run states against itself, and it matters. No probe was pre-registered. A probe states, before a search runs, what should be findable and where; only that turns an empty result into evidence of absence. Without one, none of the five blanks below counts as evidence that the thing is not there — only that this run did not find it.

Five searches came back empty. They are printed as empty.

The measured numbers

The figures that look contradictory are measuring different questions.

Two per cent. Three per cent. Ten per cent. Sixty-six per cent. Published figures on flat-earth belief appear to disagree by an order of magnitude. Read with their question wording attached, they stop disagreeing.

The 2% is Americans who “resolutely” say the Earth is flat. The 3% is Britons saying it is probably or definitely true, against 93% who say it is false. The 10% is a different instrument entirely: a composite counting agreement with any of three separate claims — flat Earth, faked Moon landings, or vaccine microchips.

The same graded structure appears inside a single poll. 84% of Americans say the Earth is round; a further 5% say they used to believe that and now have doubts; and only 66% of millennials hold the round-Earth answer firmly. A headline number indexes how much firmness the instrument demanded and how many claims it bundled together — not how much of the public disagrees.

None of this settles whether the answers are sincere, and the run says so plainly. The claim that trolling inflates these figures received no direct test in this record. What exists is indirect: a sample recruited in person at the first Flat Earth International Conference, about 600 attendees at the third, and conventions held in Brazil, Britain and Italy. Those establish that a sincere core exists. They cannot bound what share of the 2% it accounts for.

The best-supported reading

The strongest thing in this ledger, printed with the attack made on it.

One explanation collected far more supporting passages than any other: that pre-existing disposition — distrust, powerlessness, religiosity, conspiracy mentality, a hunger for meaning — underlies the belief. Sixteen passages leaned toward that explanation. Then the run attacked the finding drawn from it.

The supporting material is real. In one survey, flat-earthers were more than twice as likely as the general public to call themselves “very religious”: 52% against 20%. Measured against a national sample matched on religiosity and belief in evolution, they scored significantly higher on conspiracy mentality. They also described themselves as more skeptical and more logical than that national sample.

All of it is cross-sectional and correlational. It establishes that these traits travel with the belief. It does not establish that they came first, and the finding is typed as inferred rather than observed for exactly that reason.

The run's own stress test then names the weakness: the profile leans on a narrow set of overlapping review citations and small convenience samples drawn from conference attendees, which may overstate how widespread conspiratorial mentality and religious literalism are among believers generally. The condition under which the finding fails is recorded beside it, in advance, so a reader can watch for it.

The open question

The study that would settle it has not been done.

The dispute between disposition and platform has an obvious decisive test: does algorithmic exposure predict new belief after controlling for institutional distrust, epistemic style and identity? No evidence in this record addresses it.

The platform side reaches only as far as mechanism plausibility. One reviewed paper characterises YouTube visual arguments and Twitter written arguments as having a strong impact on belief, but reports no controls, no randomisation and no longitudinal design. YouTube's own 2019 announcement singles out “claiming the earth is flat” as content its recommender would begin demoting — which names a mechanism, but is a corporate policy statement, not a measurement of anyone's belief changing.

The defensible reading is a two-stage account that neither explanation as written captures: disposition selects who is recruitable, platforms determine who is reached. That reading was itself flagged for attack, because it assumes exposure cannot generate conspiratorial thinking in someone who had none — which nobody has tested either.

This is the honest shape of the answer. Not a verdict between the two, but a named, runnable experiment that nobody in this record has run.

Mechanism plausibility is the highest causal criterion the platform side of this ledger reaches.

The score

It scored itself 14 out of 100, and printed the arithmetic.

Eight findings were asserted. Every one of them computed to a confidence of unknown. The overall evidence score came to 14 out of 100, and the four numbers it was made from are printed beside it.

Corroboration scored zero because zero of fifteen source snapshots could be traced to an upstream origin, and provenance scored zero for the same reason. Unknown provenance is unknown independence: four hundred pages paraphrasing one blog post are one source, and a source whose origin was never established cannot be told apart from an original one.

Falsification scored 17 of 20 — the counter-search genuinely ran, and genuinely returned things. Saturation scored zero: the second retrieval round found nine source families, of which seven were new, so the search was nowhere near exhausted when it stopped.

The score is not a probability, and the run says so. There are no calibrated priors and no likelihood model here, and an ordinal matrix of pluses and minuses cannot honestly be converted into a percentage. A reader who rejects the weighting still has every count it was made from.