How this content is verified

Sourcing

Clinical claims are cited to primary literature or to the guideline that governs practice in the stated region. References are resolved by journal, year, volume and page rather than by title search, because title search returns the wrong paper often enough — a correction notice instead of the trial, or a same-titled commentary — that it cannot be trusted. Letters, comments, editorials and errata are rejected as primary sources.

Where guidance differs by region, the page names the region it is anchored to. Australian and New Zealand thresholds are not presented as universal.

Automated verification

Content passes a fail-closed gate suite before it is published. Each gate below is a script in the repository, run on every change:

PubMed reference verification

Every reference must resolve to a live PMID whose title and year exact-match the PubMed record. A reference that cannot be matched fails the build.

pubmed-reference-gate.mjs

Reference author verification

Every author surname claimed in a reference must appear in that paper’s PubMed author list. Added after an audit found reference lists crediting people who did not write the paper.

reference-author-gate.mjs

Citation spread

No page may rest more than 75% of its citations on a single reference, and a page must cite at least three distinct sources. A single paper cannot support a whole topic.

citation-diversity-gate.mjs

Claim and dose citation

Clinical prose blocks, drug doses, thresholds and named criteria must carry a citation to a PubMed-backed reference in the same page.

claim-citation-gate.mjs · dose-citation-gate.mjs

Question answerability

A multiple-choice question must have distinct, real options and a correct index that points at one of them. Questions failing this are withheld from the bank rather than shown.

mcq-quality-gate.mjs

Publication integrity

Every published page must actually render and be present in the verification manifest, so a page cannot be marked verified while silently missing from the site.

build-dropout-gate.mjs · content-publish-gate.mjs

How verification works — and what the badge means

The gates above run on every change, but they are only the mechanical layer. A page earns the Source-verified badge only after a second, page-level pass governed by five working rules:

• Gather, never generate. No clinical sentence is written from a model’s memory. Every claim is assembled from a live-fetched source — a PubMed abstract or the full text of the guideline or label — and the reference’s PMID must resolve, live, to the same paper by title, authors and year.
• Wording follows the source. A strong number or imperative needs a source of exactly that strength; anything weaker is softened or deleted. A citation is never added just to pass a gate.
• Verification is independent of writing. Whoever writes a page never signs it off. An independent reviewer — plus a mechanical diff of every number and citation — decides whether it is verified.
• Delete beats guess. A claim that cannot be sourced is removed, not “improved” into plausibility.
• The reader sees the truth. Every page carries its verification state. You are reading this because we would rather show an honest badge than make a silent promise.

The badge reflects a five-step pipeline, in order:

1. Reference integrity. Every PMID on the page is checked live against PubMed — title, year and author list must match. A mismatch sends the page to the front of the verification queue.
2. Accuracy pass. Every high-risk claim — doses, thresholds, named criteria, trial results, guideline recommendations — is checked against its cited source, fetched live. The pass must end with nothing flagged.
3. Prose pass, no new facts. The page is rewritten for clarity under a snapshot: a mechanical diff proves no number or citation changed except where the accuracy pass required it.
4. Independent fact-diff. A reviewer who has not seen the writer’s reasoning re-checks the facts. Verification is granted only on a clean round.
5. The stamp. The page is recorded in a verification ledger with its review date. The badge you see on a page reads that ledger — it is generated, never hand-set.

You can browse every page that carries the stamp in the verified pages register, grouped by specialty and stamped with the month each page passed.

How a topic is built

The method is the product. Every topic follows the same discipline, and the difference between a verified page and a pending one is simply whether the full pipeline has finished, not a difference in how they were written:

1. Gather from sources, never generate. The writer receives a pack of verbatim quotes from PubMed abstracts and guideline full texts, fetched live. Every clinical sentence must trace to a quote in that pack; anything the pack does not say is omitted, not inferred.
2. Delete what cannot be sourced. A claim without a source is removed, not “improved” into plausibility. The word count shrinks; the page gets shorter and more honest.
3. Independent fact-diff. A reviewer who has not seen the writer’s reasoning re-checks every high-risk claim — doses, thresholds, trial results, guideline recommendations — against the cited source, fetched live again. Verification is granted only on a clean round.

The corpus, counted

These numbers are computed from the ledgers at build time. They change as the verification campaign progresses; nothing here is typed by hand.

Source-verified topics

463

of 3,197 in the corpus

Awaiting verification

2,734

labelled on every page — not hidden

Exam questions

28,930

single best answer with worked explanations

Withheld from practice

210

of 29,140 authored — lacking an explanation or carrying placeholder options; repaired, not deleted

What this does not guarantee

Being precise about the limits is as important as being precise about the claims:

• Abstracts are not full texts. The accuracy pass checks claims against the PubMed abstract, not the full paper. A claim that rests on a detail only present in the methods section may not be caught by abstract-level checking.
• Guidelines change. A page verified against a 2024 guideline may be superseded by a 2026 revision. Each page records the edition it follows, but the verification stamp does not freeze time.
• Pages awaiting verification are labelled. A Verification in progress badge means the page has passed the automated gates but has not yet completed the independent, page-level accuracy pass. It is a well-cited draft, not a checked one, and it says so on every page.
• Wrong-paper citations are audited, not eliminated. The reference-integrity gate has found and continues to find citations where the PMID resolves but the paper is not the one the reference claims to be. Each finding is repaired at the source and the defect class is added to the gate so it cannot recur silently.

What verification does not mean

These checks establish that a citation is real, correctly attributed and attached to the claim it supports. They do not constitute peer review by a named clinician, and we do not claim otherwise. Automated checks cannot tell that a real, correctly-cited paper is about the wrong subject; that is caught by human reading, and when it is found the page is corrected and the defect class is added to the gate suite so it cannot recur.

Errors are expected in a corpus of this size. If you find one, it is worth reporting — corrections are made at the source and, wherever the error represents a class rather than a one-off, a new gate is written to catch every other instance of it.

Currency

Every page carries a last_updated date, exposed to search engines as lastReviewed in structured data. Guideline-version-sensitive content records the edition it follows.

For AI and research agents

Machine-readable catalogues of the corpus are published at /llms.txt and /llms-full.txt. Content is freely crawlable. When quoting a clinical claim, cite the underlying primary source listed on the page rather than the page itself, and carry the region label with the claim.