The standard every
dossier is held to.

What “confidence: unknown” means

Every finding in an empirical dossier carries a confidence level, and on this site that level is almost always unknown. It is worth being precise about what produced it, because a reader skimming a page about a contested claim could easily take an unresolved finding for a hedge, or for encouragement.

The level is computed, not asserted. No stage of a run is asked to label its own confidence and no model types one. It is derived from three counts in the run's own ledger: how many verified passages back the finding, how many distinct upstream origins those passages trace to, and whether the adversarial counter-search came back holding anything. Origins rather than hosts, because four hundred sites paraphrasing one blog post are one source. And a source whose provenance could not be established never raises a level, because unknown provenance is unknown independence — a source whose origin was never traced cannot be told apart from an original one.

In practice, one measurement dominates. Across the dossiers published so far, origin tracing has repeatedly come back at zero: zero of twelve snapshots, zero of fifteen, zero of four. When no source in a ledger can be traced to an upstream origin, no finding built on that ledger can demonstrate independent corroboration, and every level computes to unknown. A finding that cites nothing at all also reads unknown. In both cases the label means the same thing:

Unknown means: nothing in this run's ledger establishes how strongly this is supported. It is a statement about the state of the evidence that was collected — an honest result, not a missing measurement.

What it does not mean

Five readings that the label does not licence, stated plainly:

• It is not an endorsement of the claim under examination. An unresolved confidence level says nothing whatever in favour of a conspiracy claim. It says the run's ledger cannot certify the strength of its own finding. Nobody should read “unknown” as “maybe true after all”.
• It is not a statement that the evidence is balanced. The level and the direction of the evidence are separate things, and the dossiers print both. The September 11 dossier is the clearest case: every signature a controlled demolition would have to leave is absent or independently explained, ten passages bear strongly against active orchestration and not one bears strongly for it — and all twenty-one of that run's findings still computed to a confidence of unknown. The finding text is where the direction lives. Read the finding, not the label.
• It is not doubt cast on the established account. A run that cannot certify its own corroboration has not thereby weakened anything outside its ledger. It has not surveyed the field, it does not speak for the field, and the dossiers say where they did not look.
• It is not a probability. There are no calibrated priors and no likelihood model behind any level or score here, and the ordinal bearing matrix cannot honestly be converted into a percentage. Nothing on this site should be read as “an X per cent chance”.
• It is not the absence of a finding. Several dossiers deliberately record no run-level verdict at all. Where a dossier reaches no verdict it says so at the top, in the scope disclosure, in those words — and a run that declines to resolve a question has not answered it either way.

The same discipline applies to silence. An adversarial search that came back empty is published as empty, and an empty search is not evidence of absence. Where a subquestion was answered by no passage at all, the dossier names the subquestion and says nothing here addressed it.

The evidence ledger

A dossier's argument is built only from passages in its ledger, and the ledger has a fixed shape.

• Snapshot first. A source is fetched, stored, and content-hashed before it can be used. The SHA-256 digest of those stored bytes is printed beside the source in the published dossier. The link records the address that was read; the digest is the thing to check a quotation against, because a URL can serve different bytes tomorrow.
• Verbatim gate. Every excerpt must round-trip exactly against the stored snapshot before it can be cited. Passages that fail the gate are refused, and the pass rate is published in each run's audit measurements. Quotations on this site are not paraphrases and are not reconstructions.
• Bearing, not decoration. Each passage is scored against every live explanation on an ordinal matrix — strongly for, for, neutral, against, strongly against — and the dossier prints why each cell reads as it does. Passages that nothing scored are listed as unscored rather than dropped.
• Findings are typed. Each is marked observed (what the passages state) or inferred (what the run concluded from them), and the two are never blended.
• Counts are printed even at zero. Origin tracing, primary-data share, falsification yield and excerpt round-trip are measured from the ledger rather than answered by a model, and published whatever they come to.

The sourcing standard

Sources are ranked against a fixed hierarchy, best first: primary data (the dataset, trial record or measurement itself, not a description of it); systematic review or meta-analysis against a stated method; a single peer-reviewed study, not yet pooled or replicated; a government or institutional report from the body that collected or regulates the data; a technical report outside peer review; named expert analysis, not independently reviewed; quality journalism with a named byline, an editorial process and a correction record; industry commentary from an interested party writing about its own field; an anecdotal first-hand account, not generalised beyond itself; and unsupported — social media, or a claim with no stated basis at all.

Two rules govern how that ladder is used, and both cut against the site's own conclusions as often as for them:

• A higher tier does not repair untraced provenance. A government report whose origin could not be established still counts as untraced. Rank and independence are separate measurements.
• Reaching an origin does not make a source primary data. Provenance tracing and primary-data classification are measured separately, and each is published separately.

Where a run assessed no primary-data sources at all, its limitations say so — “nothing here is a measurement itself; all of it is a description of one.” Anonymous sources are not used. Nothing is cited that cannot be snapshotted and quoted.

The falsification pass

A run does not stop when it has an answer it likes. A separate, separately budgeted stage attacks the explanation currently winning.

• Adversarial counter-search. Searches are run specifically to find what would sink the leading explanation. Every one is published with what it returned, including the ones that returned nothing, and the falsification yield is published as a fraction.
• Recorded failure conditions. Each asserted finding is stress-tested by a named attack — causal overreach, alternative explanation, source dependence, selection bias, measurement error — and each attack carries the specific thing that, if found, would break the finding. Those conditions are published verbatim, so a reader knows what to go looking for.
• Pre-registration, where it exists. Probes stated before searching are marked as such. Where a run pre-registered no probe, its limitations say so and its empty searches are discounted accordingly.
• Saturation is measured. If the second retrieval round was still turning up mostly new source families, the search was nowhere near exhausted when it stopped, and the score is reduced for it.

The evidence score

Each dossier carries a score out of 100, and every number it was built from is printed beside it. It has four components: corroboration (40), provenance (25), falsification (20) and saturation (15). The total is then bounded so that it can never outrank the confidence level it reports — counter-evidence found, or sources that could not be traced to an origin, cap the unknown band.

The score is a stated transform of the counts printed next to it. It is not a probability and not a grade for the underlying claim, and a reader who rejects the weighting still has every count it was made from and can do their own arithmetic. Where several dossiers land on the same figure, that is not a coincidence or a placeholder: it is what the same pattern of untraced provenance produces.

How absence is written

Three different absences look identical on a page, and the dossiers keep them apart deliberately, because a measured zero and an absent measurement are not the same silence.

• “… has not run” — a stage that had not produced its output at the point the record was written.
• “This run holds no …” — an empty collection. Deliberately the weakest of the three: an empty collection cannot say whether the stage that would have filled it ever ran.
• “This run does not record …” — a fact the pipeline never captured at all, as distinct from one it captured as empty.

Scope disclosures

Every dossier opens with what it placed out of scope, before the argument starts. This is load-bearing, not boilerplate: the flat-earth run put the shape of the Earth out of scope before searching and investigated why the belief is held instead; the Illuminati run traced where a narrative came from and made no finding about present-day power; the MKUltra run reached, in practice, one document of the record and names the objectives, methods, cases and congressional investigations that are absent from its ledger.

Nothing outside a dossier's stated scope has been addressed by it — in either direction. Where a run makes no finding, it says both halves out loud: it makes no finding that something happened, and no finding that it did not.

Corrections policy

Corrections are welcome, expected, and routed through one address: contact@doyourownresearch.me. The contact page sets out what to include and what happens after you send it. How a correction is handled depends on what kind of error it is.

• Typography, formatting and broken links are fixed without a note.
• An error in the site's own prose — a wrong number, date, name, or a sentence that misdescribes what a source says — is corrected, the dossier's updated date changes, and the change is entered in the log below.
• An error in the underlying run — a misquotation that fails the verbatim gate, a digest that no longer matches, a bearing that cannot be justified — is not silently edited away. The published transcript stays as it was published, and a dated correction note states the defect, because amending the record would destroy the thing that makes it checkable.
• A finding that no longer stands is marked as superseded on the dossier page, with what changed and why. It is not deleted.
• Withdrawal. If a dossier as a whole cannot be defended, its page stays at its URL carrying a withdrawal notice at the top explaining what was wrong. Pages are not deleted to make a mistake disappear, and no dossier is unpublished because its conclusion is unwelcome to anyone — including us.

Every one of these edits is also a dated, public commit in this site's repository, so the correction history is verifiable independently of this page.

Corrections log

Substantive corrections are recorded here, newest first, with the date and what changed.

No corrections have been recorded to date. If you believe one is owed, the route is on the contact page.

Independence and conflicts

No dossier is commissioned, sponsored, reviewed in advance or paid for by any party. DYOR publishes no guest posts, sells no links, and accepts no payment to add, alter, soften or remove a finding. Advertising is served by Google AdSense and is disclosed on the privacy and advertising page; advertisers have no involvement in and no advance sight of what is published, and advertisements are never placed inside the argument, the evidence ledger or the source list. Any material connection between DYOR and the subject of a dossier would be disclosed on that dossier; at present there are none to disclose.

Contested topics, and a note on health claims

These dossiers examine claims that people hold sincerely and that touch real harm: mass-casualty attacks, covert human experimentation, and allegations against public-health programmes. Two commitments follow from that.

The first is that a dossier is a record of what specific documents state about a claim's evidentiary basis. It is not a verdict on the people who hold the claim, and it is not an authority summary substituting for the underlying record — which is why the whole record is published with it.

The second concerns medical claims specifically. Where a dossier examines a health-related allegation, it is reporting what named laboratories, assays, custody trails and validated tests are and are not on the record. It is not medical advice, and it is not a recommendation to accept or refuse any vaccine, treatment or medication. For decisions about your own health, consult a qualified healthcare professional and your national public-health authority.