Every factual claim in this publication is tied to a source, and the tie is recorded rather than implied. This article describes how, because a reader who wants to check something should not have to reverse-engineer the method first.
Sources are objects, not links
A markdown link at the end of a sentence records a URL. It does not record what kind of thing the source is, when anybody read it, whether it has since been replaced, or which sentence it actually supports.
Each source here is its own record, carrying its title, publisher, type, jurisdiction where relevant, publication and effective dates, the day a person read it, how they reached it, whether anything supersedes it, and a note in that person's own words about what it says. All of them are listed on the sources page.
The retrieval note is the field that does the most work, and it exists because of a specific failure: a successful HTTP response proves that a server answered. It does not prove that anybody opened the page, and it certainly does not prove that the page still says what a citation claims it says. Reading a statute through a private consolidator is a different fact from reading it in the official register, and the note is where that difference is recorded rather than lost.
Not all sources are the same weight
Sources are ranked, strongest first: legislation and official regulatory text; regulator publications; official registers; recognised clinical guidelines; systematic reviews; peer-reviewed studies; institutional summaries; professional body positions; journalism; company statements.
The ranking is not decorative. An article carrying high or critical risk may rest only on sources at or above peer-reviewed research, and the build refuses one that rests on anything weaker. Institutional summaries are readable and usually correct, and every one of them is somebody else's account of a source nobody here has read.
Company statements sit at the bottom, and that includes GMedAC's own. An article about GMedAC citing GMedAC is a company statement however carefully it is written, which is why the articles here about the platform are not published at high risk.
A source list also has to be used. A declared source that no claim rests on fails the check, because a list of references at the foot of an article is how a piece looks evidenced without being evidenced.
What happens when this is wrong
Something published here will be wrong. The question is what happens next.
A correction is its own record. It quotes what the article used to say — the actual sentence, not a description of it — alongside what it says now and why it changed. Corrections are listed on their own page and are never removed.
A correction that changes a medical or legal conclusion is marked as material, and two things then follow automatically: the article's modified date has to move, so every cache, feed and search engine sees a changed document rather than the same one; and a reviewer has to sign it, not only an editor. Neither is optional, and the build refuses a material correction missing either.
That model is borrowed from how GMedAC's own legal research is kept. Its source register records thirteen statements corrected during a single research pass, each preserved beside the original rather than replacing it, on the principle that the record of what was believed and when is the thing worth keeping.
One thing this publication does that the register does not: record how each
source was reached, per source. The register describes its retrieval routes once
for the whole package, and says so itself — a limitation added, in its own
words, because an earlier draft promised more than it delivered. A reader of a
single row there cannot tell which of three official routes the text came from,
or whether it came from a private consolidator instead. That gap is why
retrievalRoute is a required field here rather than a paragraph somewhere.
Content goes stale, and it goes stale differently
Every article carries the date its facts were last checked and the date they are due to be checked again.
What happens when that date passes depends on what being wrong would cost. High-risk content past its date stops the build; it cannot stay published. Anything carrying emergency information stops the build regardless of its risk level, because a stale emergency number is wrong at every level somebody happened to assign, and the reader who needs it is the least able to notice.
Ordinary explainers past their date currently produce nothing. That is a gap rather than a lighter treatment, and it is written down as one: the reasoning for not failing the build is sound — a build that fails because a general explainer is a week overdue teaches people to move the date rather than read the article, and a control people route around is worse than no control — but the reporting that was supposed to replace it has not been built. Saying so is more useful than describing a queue nobody maintains.