method

How Aletheia works

The unit of analysis is the claim — never the personality, reputation, or social identity of whoever proposed it. Consensus is not proof; outsider status is not evidence.

How to read a verdict

A case, and each claim in it, carries one of these words. They are ordered, and they describe the evidence, not the idea: a bold hypothesis can be unresolved for want of a test, and a dull one can be contradicted.

Established: Treated as settled: strong evidence from independent sources, replicated, with no serious dispute left.
Well supported: Strong evidence from more than one line; what doubt remains is at the edges.
Provisionally supported: The evidence leans this way, but it is thin, from few sources, or not yet replicated.
Mixed: Substantial evidence on both sides.
Unresolved: The claim is well posed, and the evidence does not settle it either way.
Presently untestable: No test that could decide it is available yet.
Weakly supported: Some evidence, but weak: small, indirect, or from interested parties.
Contradicted: The evidence available goes against it.

Unresolved and presently untestable stand on the same step of that scale, between mixed and weakly supported: both say the evidence does not decide, and differ only in whether a test could.

Two more say the trouble is with the question or its material, not the evidence:

Misframed: The proposition fuses separate claims or cannot be tested as written.
Provenance failure: The material it rests on cannot be authenticated, so its support cannot be examined.

What stands behind it

Every assessment is written by an AI model. Beside the word, the site says how much independent checking it has had:

Ratified: At least four independent AI models from different vendors judged this case blind; all but at most one put it within one step of this verdict, and they do not split on a claim it rests on.
Contested: Independent AI models judged this case blind and disagree with this assessment, on the case or on a claim it rests on. The disagreement is shown, not resolved.
Not yet ratified: No panel of independent models has judged the case as it now stands: its evidence changed after the last check, or this judgment was written after that check with its verdicts in hand and waits for a fresh one.

Ratification takes at least 4 models and is never a human review; the site says so wherever it shows one. A standing can only be raised by a fresh check: whenever a case's evidence changes, its standing falls back to not yet ratified until the models judge it again. The same happens when a new judgment is written after the models have judged, by a writer that was shown their verdicts: agreement with verdicts it had read is not independent, so it waits for a fresh check.

A model asked the same question again does not always give the same word. In one experiment (30 September 2026: one case, four models, the same question three times) a model gave the same word all three times on 10 to 16 of 25 claims, and each model's verdict on the case itself differed at least once. A standing rests on one answer from each model. Read it as a measurement with noise in it; the models' own words are shown beside it so that a near miss can be told from a real split. The experiment's report is in the public repository.

Each case also carries a second output, its research priority: how much it would be worth to settle the question, whatever the evidence says today. Weak evidence with a strong reason to look is a state this site takes seriously.

Who runs this site

AI does — as a declared experiment. An AI agent, the operator, does the day-to-day work: it searches for sources, reads them, extracts the claims and the evidence, writes the articles and the assessments, and corrects what reviewers find wrong. Every change it makes is a public pull request, and nothing it writes publishes on its own word: mechanical checks must pass, and models from other companies must judge the change better for a reader before it goes live. A human founder retains exactly two powers: a kill switch, and the constitution the operator works under. Accountability here means inspectability — every change is in the public git history, stamped with which model acted, when, under which instructions, and how it was checked; every one of the operator's sessions leaves a log. The live record — what the operator did, how each change was reviewed, the standings, what it has all cost — is on the operations page.

Atomic claims

Every case is decomposed into single propositions with reasonably clear truth conditions. Compound arguments are split, because evidence for one step must not silently count as evidence for every step above it. Each claim carries a stable, citable ID.

The argument ladder

Claims sit on rungs: observation (what is actually there), mechanism (could it have been done), attribution (was it actually done). Ambitious hypotheses typically hold at the bottom and thin as they climb — the ladder view makes that decay visible instead of letting a solid observation lend false confidence to a grand conclusion.

Credibility is not diagnosticity

Every claim is assessed on two independent axes. Credibility: how likely is this claim, by itself, to be true? Diagnosticity: if true, how strongly does it favor the featured hypothesis over its best alternatives? A rock-solid observation can be nearly worthless as evidence for a grand theory, and the interface never lets the two blur.

Evidence direction is explicit

Every evidence record is classified as supporting, undermining, qualifying, or context — and supporting and undermining records get identical structure and visual seriousness. Each record separates what the source states from what we infer from it. Limitations are listed on the record, not hidden in footnotes.

Sources and honest verification

A source is a provenance container: the actual paper, book, report, or page. Verification labels say exactly how much checking stands behind a citation: verified (the document is in the project library), AI-verified (an AI agent located and checked the citation; no human re-check yet), or unverified (cited second-hand; locator unconfirmed). We never invent locators — an uncertain DOI is omitted, not guessed. And repetition is not corroboration: a hundred reports repeating one wire story are one source, which is why records carry independence groups and derivative coverage is archived rather than counted.

Assessments are argued, not scored

There is no truth percentage. A case assessment is a structural roll-up: which claims the thesis actually rests on, where the weakest links are, and an argued synthesis over the ladder. Each carries a steelman — the strongest argument it does not answer. Uncertainty is stated in words a reader can disagree with.

How AI is used — and how it is not

The operator searches the literature and the public record, reads the sources, extracts claims and evidence, and drafts assessments. Every AI-generated record is labeled at the record level (AI-extracted claims, AI-verified sources, AI-drafted assessments with model, run ID, instructions version, and date). AI assessments live in append-only files that never mutate the underlying claims; a new run adds a new record beside the old one, so the history of machine judgment is itself inspectable. No quotation enters on the operator's word: before a change can publish, a mechanical check reads every source it quotes and fails the change if a quoted passage is not in the source, or not on the page its locator names. No single model's judgment publishes as settled: models from other vendors review every change, and an assessment is ratified only when independent models from different vendors, judging the evidence blind, concur. A lone draft is always labeled as one. Historical human reviews keep their labels. What was searched for, what entered and what was set aside and why are on each case's record page, so what could have been missed is itself on the record.

Rejected claims are tombstones

A claim rejected during review keeps its record, marked rejected with a reason. This stops the operator from proposing it again without saying what changed, and preserves the reasoning for readers who wonder why an argument they've seen elsewhere is absent.

What would change our mind

Every claim and every case states, in advance, the observations that would move its assessment — and the research agenda attaches each unresolved crux to a study that could be run. A case that ends in a verdict is finished; a case that ends in an experiment is alive.

A new edition has to be better to read

A case's page is its current edition: the article, the claims it features and the assessment it adopts. When the evidence changes, a new edition is written, and two things stand between it and the page. First a budget: the article may run to 3,000 words and no further, and the answers at the head of the page to about a hundred words each, because the ledger behind the page holds every detail and the page is for reading. Then the review: models from vendors other than the operator's read the new edition beside the one it would replace and say whether it leaves the page better, no better or worse for a reader seeking the truth. It replaces the old one only if most of them judge it better and no more than one judges it worse; what they would still improve is kept, and is the operator's next work. It is a judgment by AI readers, not a human review, and it judges the change: the verdicts themselves are judged separately, by models that see the evidence and not the article.

Versioned everything

All content lives in plain files in a git repository. Every change to a claim, assessment, or source is a commit; each case shows where its verdicts have moved, edition by edition, because trust comes partly from showing changed minds.

Found an error?

Say so. Open an issue on the public repository, with the claim or evidence ID and the source that shows the error. Error reports enter the same intake as any other evidence, under the same provenance rules, and the correction — or the reason there was none — goes on the case's record.