What we count as a source — and what we refuse to count
Every source captured for an episode is sorted by its publisher before a single claim is checked against it, and only the top tier lets a claim pass without a person reading it. These are the actual lists — including the sources we could have counted and deliberately do not.
- A source is tiered from its web address alone, before any claim is checked against it.
- Only a top-tier source lets a claim clear without a person reading it.
- Preprints, journals with contested peer review, and bare DOI links are deliberately kept out of that tier.
- The lists below are rendered from the ones the checker reads — not a copy maintained by hand.
These lists are one step of the check. The rest of it is on the fact-check page.
Anyone can publish a list of the sources they trust. It costs nothing to write and nothing to keep, which is exactly why it proves nothing. The list worth reading is the other one — the sources that look authoritative, that we could have counted, and that we decided not to. A preprint is a real study. A paper in a journal with contested peer review is often perfectly sound. A DOI link usually resolves somewhere reputable. None of them clears a claim here without a person reading it first, and every one of those decisions costs us review time rather than saving it.
How a source is sorted
The tier is decided from the source’s web address alone, at the moment the source is captured and before any claim is checked against it. No model is asked which tier a source belongs to, and nothing a model returns can change one: which list a host is on is a lookup, so a lookup does it. The checkers are not even told which tier a document carries — a model shown which label clears the gate has been handed a reason to prefer that document.
The classifier makes no network request, and that is why a bare DOI link cannot be top tier: only fetching it would say whether it resolves to a journal or to a preprint server, and the rule that denies the checkers web access denies it here too.
A host on none of these lists is not rejected. It is captured, checked and shown on the episode page exactly as any other source is — it simply cannot clear a claim on its own, and the claim goes to a person instead. Being absent from these lists costs a source nothing but the automatic pass, which is why they stay short rather than guessing at borderline publishers.
The lists themselves
Counted as the record itself
A claim clears without anyone reading it only when both checks agree and one of them can quote a passage supporting it from a source published here. Journals, biomedical archives and trial registries: the place the research is reported, rather than an account of it.
- pmc.ncbi.nlm.nih.gov
- pubmed.ncbi.nlm.nih.gov
- ncbi.nlm.nih.gov
- clinicaltrials.gov
- cochranelibrary.com
- nejm.org
- thelancet.com
- nature.com
- science.org
- sciencedirect.com
- link.springer.com
- springer.com
- bmj.com
- jamanetwork.com
- cell.com
- plos.org
- wiley.com
- onlinelibrary.wiley.com
- tandfonline.com
- oup.com
- academic.oup.com
- biomedcentral.com
Being on this list qualifies the publisher, not every page it serves — see “A publisher’s newsroom is not its research” below.
Whole namespaces counted the same way
Government, European institution and accredited UK academic space, matched on a label boundary, so the government namespace covers a health agency and not a lookalike domain that merely ends in the same four letters.
- .gov
- .gov.uk
- .ac.uk
- .europa.eu
The American university namespace, .edu, is deliberately absent — against the original plan, and on measured evidence: run over the source URLs of every episode in our database it produced false top-tier results, because a for-profit career college’s marketing page and a consumer-health blog both sit on .edu. Registration there is restricted to accredited institutions, which is not the same thing as restricted to research. A genuine university research host can still be added by name, where the decision is visible.
Counted as reporting on the record
Wire services, newspapers of record, science and trade press, and institutional health explainers. Real editorial standards, and often the only place a finding is readable — but an account of the research rather than the research, so a claim resting on one of these is flagged and goes to a person even when both checks agree it says what the script says it says.
- mayoclinic.org
- clevelandclinic.org
- hopkinsmedicine.org
- health.harvard.edu
- nhs.uk
- who.int
- reuters.com
- apnews.com
- bbc.com
- bbc.co.uk
- nytimes.com
- wsj.com
- ft.com
- economist.com
- theguardian.com
- washingtonpost.com
- npr.org
- scientificamerican.com
- newscientist.com
- statnews.com
- nationalgeographic.com
Short on purpose. An omission here changes nothing — an unlisted host is treated exactly the same way — so there is no reason to guess at a borderline publisher.
Trusted by people, never automatically
The list that costs us something. Each of these is a real source that a researcher would reasonably cite, and each is deliberately kept out of the tier that clears a claim on its own.
The preprint servers publish the study itself, which makes them primary in the literal sense — but not peer-reviewed, and an unreviewed preprint auto-clearing the gate on a health channel is precisely the failure this was built to prevent. The two journals run contested peer review: sound papers appear in both, so does work that would not survive review elsewhere, and a list of hostnames cannot tell them apart. The DOI resolver is not a publisher at all — the same link format resolves to a preprint server and to a top-tier journal, and only a network request could say which.
- arxiv.org
- biorxiv.org
- medrxiv.org
- frontiersin.org
- mdpi.com
- doi.org
None of these is blocked. A source here is captured, checked and cited like any other; it goes to a person before narration rather than clearing on its own. That is a review cost, not a refusal.
A publisher’s newsroom is not its research
Every publisher on the first list also runs a newsroom. Nature publishes news about research beside the research; so do the Lancet, Science, the BMJ and every government agency. A sorting rule that looked only at the hostname would hand a news write-up the same standing as the paper it is about — which is the exact confusion this whole process exists to catch.
So the address is checked as well as the host. A URL on a qualifying publisher that sits under its news, press, media, blog, opinion, comment or editorial section is demoted to the middle tier and goes to a person, and a subdomain announcing itself as one of these does the same. The demotion only ever costs a source its automatic pass — it can never promote one — so the rule can afford to be broad.
Subdomains treated as a newsroom on an otherwise qualifying publisher:
- blog
- blogs
- news
- press
- media
What these lists do not tell you
- They judge the publisher, not the paper. A journal on the first list publishes weak studies like every journal does. What the tier decides is whether a person has to look at a claim — never whether the finding behind it is true.
- They are lists of hosts, and hosts change hands and change lists. A publisher added today was not top tier yesterday, and an episode published before a change was sorted under the rules that existed then.
- They say nothing about a source with no web address. A client’s own pasted material has no host to judge, so it is never top tier — material you supply is not self-certifying, and a claim resting on it goes to a person.
- They are not a ranking. Nothing here says a wire service is worse journalism than a journal is science; it says one reports the record and the other is the record, and only the second can clear a claim unread.
Where this fits
Sorting sources is one stage of the check an episode script goes through before anyone is allowed to record it. The claims are pulled out of the script, checked twice against the text captured for that episode, and anything the two checks cannot support — including anything supported only by the middle tier above — stops the episode until a person has dealt with it.