An article that cannot say where a claim came from is an opinion in the typography of a reference work, and a reader has no way to tell the two apart. This policy exists so that every claim in this publication is traceable to a document somebody actually opened, and so that the weakest part of the chain is visible rather than smoothed over.
The evidence hierarchy
Source types are recorded in a fixed vocabulary, strongest first. The order is about what kind of authority a document carries, not about how well it is written.
LEGISLATION— the enacted text itself, in its consolidated form, from the official publisher. It is what the rule says, not what anyone reports it to say.REGULATOR_PUBLICATION— guidance or a decision from the body holding statutory power over the activity in question. It records what that body has said it expects, which is not the same thing as what the law requires, and it is cited as guidance rather than as law.OFFICIAL_REGISTER— an entry in a register maintained by the authority that grants the entitlement. It answers whether a specific person or organisation is listed, on a date.CLINICAL_GUIDELINE— a guideline from a body that publishes its method and names its panel. Its authority comes from that declared method. A guideline without one is weaker than its type suggests, and the vocabulary has no way to say so: only the reviewer can.SYSTEMATIC_REVIEW— a review with a stated search strategy and stated inclusion criteria, which is what separates it from a summary of whatever the author had read.PEER_REVIEWED_STUDY— a single study in a journal with review. It is evidence of one result under one set of conditions, and nothing broader.INSTITUTIONAL_SUMMARY— a hospital, ministry or public-health explainer. Usually accurate, usually undated, and rarely traceable back to the study it is compressing.PROFESSIONAL_BODY_POSITION— a membership organisation's stated position. The interest is legitimate and it is also present, so the position is recorded as a position.JOURNALISM— evidence that something was reported, by whom, and when. It is never the basis for a legal or clinical claim, however well sourced the reporting is.COMPANY_STATEMENT— what an organisation said about itself. GMedAC's own material sits here too, including when quoting it would be convenient.
The source record
A source record carries an identifier, its type from the vocabulary above, the document title,
the issuing body, the document's own date or version where it has one, a locator such as a URL
or DOI where one exists, dateRetrieved, and retrievalRoute. Articles reference records by
sourceIds; they do not carry loose links in prose.
retrievalRoute is prose, and it exists because a URL answers a different question. A URL says
where a document is meant to live. retrievalRoute says how this document was actually reached
and read — which listing, which section, which pages, which printed edition, which archived
copy after the original moved. Many honest routes are not URLs at all, and a field that accepted
only a URL would quietly require a false one.
A status code is not a reading
A successful HTTP status proves that a server answered. It does not prove that the document behind it contains what the citation says it contains, that it is the version the article relied on, or that any person read past the title. A link checker detects a dead link. It cannot detect a wrong one, a replaced one, or one nobody opened. A passing link check closes no source question. Sources are closed by a named reviewer stating that the document supports the claim.
Fabrication
DOIs, URLs, retrieval dates and claims of successful access are never invented, never inferred from a pattern, and never reconstructed from memory of how such an identifier usually looks. A plausible fabricated DOI is more damaging than a missing one, because it passes inspection and transfers the cost of discovery to the reader.
A source nobody opened is not a citation. Where a document could not be reached — paywalled,
withdrawn, offline, available only in a language nobody on the team reads — retrievalRoute
records that, and the claim either finds another source or drops to a basis that admits what
is actually known. Recording a failed retrieval is not an embarrassment; it is the only
alternative to inventing a successful one.
The risk threshold
An article carrying contentRisk: HIGH or CRITICAL can change what somebody does about their
health or their legal position, so it may rest only on sources at or above PEER_REVIEWED_STUDY
in the order above. Institutional summaries, professional-body positions, journalism and company
statements do not carry those articles, however sensible they sound.
The check enforces a class of document, not a quality of evidence. A weak study still satisfies
the threshold, and the only control that catches it is the reviewer named in reviewerIds, who
fails the way people fail. This is honest, not clean.
Superseded sources
Law is amended, guidelines are replaced, registers are corrected. A source that has been
superseded is not deleted and not silently swapped, because doing either erases what the article
knew when it was published. The record is kept, marked as superseded, and points at the source
that replaced it. Every claim resting on it returns to review. Where the replacement changes what
the article says, that is a correction under the corrections policy, recording previousClaim
verbatim — not an edit.
Reference list, not bibliography
A reference list contains exactly the sources that claims in the article rest on. A bibliography
contains what the author read, or wants to be seen to have read. sourceIds is a reference list.
A source declared in sourceIds and cited by no claim is refused. An uncited source lends the
article authority it did not use, and it is unfalsifiable in review, because there is no claim
against which to test whether the document says what the article implies.
A jurisdictional claim may not be sourced to GMedAC's own record of what has not yet been
researched. An entry marked EVIDENCE_MISSING in the launch matrix records the absence of
evidence, and absence is not a clearance. No check can tell that reading of the matrix from a
sound one, so this refusal is a reviewer's rather than the build's.