Deposit Standard

Editorial · Deposit Standard

The three rules everything else follows from

1. The error is almost never in the primary. It is in the transmission.

Medical evidence travels: primary study → systematic review → appraisal document → policy proposition → public argument → clinical decision. There is no verification step at any hop. Each layer trusts the layer above, applies its own transformation, and passes the result on.

Errors are not caught. They are laundered — and they arrive at the far end carrying an institutional badge.

In building one review, we verified 41 errors in the evidence chain underlying a national policy consultation. Roughly eight were in the primary research papers. The rest were introduced afterwards — in an appraisal table, in a systematic review’s one-line summary, in an author manuscript that was never corrected, in a news article, and in our own reading of all four.

Not one of them was detectable by a check that asks only whether a citation resolves. All of them were visible to someone who opened the source.

2. A check that is built and not read is not a check.

One version of an Evidence Review was deposited with thirty-four citations resolving to the wrong papers. Nine references were misaligned in the list. A numerical integrity check passed cleanly — no orphaned references, no dangling citations — because every number still resolved to a reference. It simply resolved to the wrong one.

A claim-to-reference map had been built specifically to catch this. It had recorded the misalignment in plain sight: eleven claims about a testosterone evidence review, filed under a written ministerial statement.

Nobody read it.

The map is now a gate, and the gate is a human being reading it line by line. No script can close it, because the script would be checking numbers again.

3. The criterion is fidelity to the source, not the direction of the result.

Corrections are kept because they are faithful to the evidence — never because of which way they cut.

This is not a posture. In one review, correcting an error restored an adverse safety signal that an earlier draft had wrongly dissolved. Another correction removed a claim of harm that the primary paper did not support. Some corrections cut both ways and cannot be scored at all — and where that is true, we say so and decline to score them.

We publish no tally of which way our corrections fall. An earlier version of one correction record did, and the count was not derived — it was asserted, and it was wrong. That is itself recorded as a correction. A record in which every error points the same way is a record nobody should believe.

Provenance of sources

Printed in full in every deposit:

And it applies from ER-015 onward — not backwards. This convention was adopted on 13 July 2026. Reviews published before it were produced without a per-reference provenance record, and that record cannot now be reconstructed: some sources were read in full, some in published abstract, some through a publisher’s web page. Those reviews therefore make no per-reference provenance claim, they say so in their own deposits, and nothing should be read into the silence of a reference in them. We would rather tell you which of our documents can carry this claim than let a standard adopted today appear to vouch for work done before it.

Silence is an assertion, not an absence. It says: read in full, in the original. That is a strong claim, deliberately. It is the claim the whole corpus is worth nothing without.

Every exception is stated on the reference itself, so that it can be counted. Reviews in the current house form go further: every reference carries a provenance record, including those read in full, drawn from one closed set:

Earlier reviews, from ER-015, state their exceptions in fixed sentences instead — Read in published abstract; full text not obtained. · Read in the official translation. · Not obtained. — and are silent on sources read in full.

The counts in the Verification statement are derived from the document at build time, never typed. Nothing is stored twice, so nothing can drift.

And one distinction that took us longer than it should have

“Read in primary” is not the same as “quoted from the primary.” A paper can be opened, read in full, and still quoted from its abstract — because the abstract is nearer to hand. We did exactly this: a set of bone-density figures was taken from a paper’s abstract, in a paper we had read in full, and the abstract disagreed with the paper’s own Results at three of four measurement sites.

This failure passes every provenance check, because the provenance claim is true. The rule now reads: figures come from the Results. Where an abstract and a results section disagree, the results text governs and the disagreement is stated.

How certainty is marked

A single certainty rating cannot carry several different questions, and collapsing them is how most public argument about medical evidence goes wrong. Our graded reviews answer them separately, on these axes:

An axis that does not apply to an outcome is left out, never marked “not applicable”. An axis that applies but cannot be answered is never left out: it is marked with the reason it cannot be answered. Each review states once, at the start, which axes it uses.

These come apart constantly. “The evidence is weak” usually means causal attribution is weak — which says nothing about whether the effect is large, or whether it matters.

The marks

Every mark is followed by its reason. A mark with nothing after it is not a grade.

A note on our own inconsistency. Reviews published before ER-015 use a single combined traffic light, described on Reading the Evidence, which folds these questions into one mark. Folding them is the thing this scheme exists to stop. Those reviews stand as published. Each moves to this scheme at its next substantive version, if one is produced; no version is produced only to convert it. Until then, the corpus carries both, and you should know which you are looking at.

Certainty marks are editorial judgements against a stated scheme. They are not formal GRADE ratings, and they are open to challenge.

What stops a deposit

These are enforced in code. The build halts; it does not warn.

And one gate that cannot be automated: the claim-to-reference map must be read, line by line, before deposit. Not merely built. Building it and not reading it is how thirty-four citations pointed at the wrong papers through a check that passed.

The archive gate

Every reference claiming to have been read in full has the document in a source archive, and the build refuses to run without it. The archive is keyed by DOI, and each source is checked weekly against PubMed for corrections, errata and retractions.

Until 14 July 2026 this section said the opposite. It said the gate was off, that the claim a source had been read in full rested on our word rather than on an artefact, and that we would rather tell you so than describe a check we were not running. It is now running.

Switching it on found five things in a review that had already been deposited four times. One reference was superseded — two corrigenda had been issued, and the article was never amended in place. One had been corrected in 2022. One could not be obtained at all: the claim it carried was true, and it was removed anyway, because a claim that cannot be evidenced is a claim that cannot be made. One passage carried two citations, neither of which supported it — a claim with two irrelevant citations is worse than a claim with none. And one archived source had a colon where its identifier needed a slash: a wrong identifier does not fail. It answers.

None of it was found by reading more carefully. All of it was found by a build that refuses.

Corrections and versions

Every Evidence Review carries a correction record, published in full. Three columns: what was wrong · where it came from · what changed. It names our own errors, and it names their causes.

It is not an apology. It is the credential. A review published under its author’s first name only, with no correction record, is asking to be trusted. One with a correction record is asking to be checked. (Corrected, 14 October 2026: this said “a pseudonymous review”. Eden is the author’s legal first name; her surname is not published. Why.)

Every version receives its own permanent identifier (a DOI) under a single concept DOI for the review. Superseded versions remain accessible.

We do not silently correct. A corpus whose corrections are invisible has no credential at all.

When we find an error in someone else’s work

We find errors in published papers, in systematic reviews, and in the appraisal documents that inform national policy. What we do about them is governed by four rules.

1. Verification means the document was opened. An error is not an error until we have read the source and checked the claim against it. Anything unverified is published nowhere and reported to no one.
2. Report first. Publish afterwards. Errors go to the body that can correct them, in writing, before they go into anything we publish. A finding that has been reported and left unaddressed is a record. The same finding published without warning is an allegation.
3. One notice, once, everything verified. A drip of corrections is a campaign.
4. Trivial errors go to the people who can fix them, not to the people who would enjoy them.

Where a review reports errors in a third-party document, the deposit states plainly: these discrepancies were reported to [body] on [date]; as of [date] the published documents are unchanged. That single sentence is what a scientist does and a campaigner does not.

Every finding those four rules govern is now recorded in one place. The Evidence Transmission Register lists each error found in a published document — what the document said, what the source said, the date it was reported, and what happened next — and it lists our own errors in the same list, on the same terms, in the same styling. It is what makes rule 2 checkable rather than asserted: it shows the interval between a report and a correction, including the intervals that are still open.

It also records a failure of this page. On 13 July 2026, the day this standard was first published, we sent a notice alleging errors in an article that had quoted its source correctly. We had compared the figures against a paper’s abstract without holding the paper. The rule that forbids precisely that — verification means the document was opened — was published here on the day it was broken. The notice was withdrawn in full on 28 September 2026, seventy-seven days later. The notice, the withdrawal and the interval are all in the register.

What this standard does not do

• It does not make our reviews peer-reviewed. They are not. They are narrative reviews, written by one author, under adversarial review. They do not follow PRISMA and they do not use a formal risk-of-bias instrument.
• It does not make them right. It makes them checkable, and it makes their errors visible when they are found. Those are different things, and the second is the only one we can promise.
• It does not replace clinical judgement. Nothing we publish is a prescribing protocol or a policy position, and no clinical decision should rest on it alone.

Version history of this standard