How we source, check, and correct every page

What counts as a source here?

Two kinds of document, and nothing else: the primary literature, and documents published by a regulator. Everything printed on this site as a fact has to trace back to one of those two, by name, with a link.

Primary literature: the paper itself, carrying a DOI that resolves or a PubMed ID that exists. Not a summary of the paper, not a news article about the summary, and not a post quoting the news article.

Regulator documents: an approved product label on DailyMed, a document from the FDA, a registry entry on ClinicalTrials.gov, or a listing on the WADA Prohibited List. These carry weight because anyone can open them and read the same words we did.

The test is identical either way: the source is named in the text, linked in the references, and checkable without asking us for anything. If you cannot get from a number on the page to the document it came from in one click, we have failed the standard.

What does not count: a secondary summary, a vendor's own marketing copy, a certificate photographed and reposted, or a figure that circulates widely with no traceable origin. When we cannot find the origin, the number does not go on the page at all.

How does a figure get matched to its source?

The rule is same arm, same timepoint, same population: a number lifted out of a trial carries the trial's conditions with it. A result measured in one group at one interval does not become a general figure by being repeated without them.

We do not merge trials to build a comparison: two studies that measured a similar thing under different protocols cannot be set side by side as though one experiment produced both columns. Where a head-to-head trial exists, we use it and name it. Where it does not, we report each study separately and say plainly that no direct comparison was run.

Figures the site computes are marked as ours: a conversion or a ratio worked out on the page is not a finding from a study, and it never gets dressed as one. The arithmetic is printed alongside it so the result can be re-derived rather than trusted, which is the same reason the dose math section shows its working line by line.

What language will you not find on a page?

Borrowed authority is banned outright: the stock phrasings that gesture at unnamed studies, unnamed research, or unnamed experts are on a blocked list, and a script greps every page for them before it can ship. Each one is a way of sounding sourced without being sourced, and each one is exactly what a reader cannot check.

Unsourced superlatives go the same way: the strongest, the safest, the most effective. If a ranking is real, a document says so and we cite the document. If no document says so, the sentence comes out.

The rule is subtraction, not softening: when a claim cannot be tied to a real source, we delete the claim. We do not hedge it into a maybe and leave it standing, because a hedged unsourced claim is still an unsourced claim doing work on the page.

We never invent a citation: no half-remembered author, no plausible-looking DOI, no reference that was never opened. Every reference on this site was pulled and read before it was printed, and inventing one would be the single worst thing we could do to a reader here.

Why does no page here pick a dose?

Because that decision belongs to a clinician with your labs in front of them: we have no access to your bloodwork, your history, or your prescriptions, and no amount of published data closes that gap from a web page.

Numbers appear in exactly three shapes: what a named study tested, what a named study reported, and what an approved label specifies. None of the three is aimed at the reader, and none of them is a suggestion about what anyone ought to do.

The calculators convert, and nothing more: the units to mg converter and its siblings take whatever figures get typed into them and return the arithmetic. They propose nothing, store nothing, and send nothing anywhere.

The short version: we report, we do not prescribe. A page that told you what to put in a syringe would be practicing medicine badly, at scale, on strangers.

Who checks a page before it ships?

The load-bearing rule is that the writer never certifies its own page: a draft that grades itself has measured its own intent, not its accuracy. Every check that decides whether a page ships is run by something other than the thing that wrote it.

First, the sources get assembled: before a word is drafted, the papers and labels are pulled and each figure is located inside them, down to the table or section it sits in. A fact that is not in that pack cannot reach the page, which is what makes the later checks mechanical instead of a re-reading of the internet.

Then the page is drafted against that pack: the builder fills the frozen page template with the assembled material and emits a claim map beside it, one line per attributed number, pointing at the source and the place inside it.

Then the scripted gate runs, and it fails closed: the checks in the next section run against the file itself. Zero failures or the page does not ship. There is no override for a page that is nearly clean.

Then an independent citation audit runs: a different model than the one that drafted the page re-checks every reference and every figure against the sources, without the drafter's notes to lean on. Separating the two is the point: a writer auditing its own citations mostly confirms what it already believed.

Then the findings get reconciled: every issue raised is reproduced or killed against the actual file before anything is changed, so a confident-sounding false alarm cannot rewrite a correct page. Confirmed findings go back to the builder to fix, and the auditor checks the fix.

And every confirmed finding becomes a new script check: when a real bug is found once, a check for that whole class is added to the gate in the same pass, so the next page is caught by a machine rather than by another review. The audits are supposed to shrink over time, not repeat.

Alongside all of that: the disclaimer at the foot of every page states that pages are reviewed by medical professionals before they ship. We publish no individual reviewer's name and no credential we cannot show you, and the byline stays the team, deliberately.

What does the scripted gate actually check?

It is deliberately dumb and completely unforgiving: it cannot judge whether a study was any good, which is why the citation audit exists. What it can do is refuse to let a known failure class out of the building, every time, for free.

• Dashes: em dashes and en dashes are counted, and the count has to be zero. House style, checked rather than trusted.
• Banned vocabulary: the borrowed-authority phrasings above, plus a denylist of strings that must never appear on this site for reasons of its own.
• Dose-instruction patterns: any sentence shaped as an instruction to the reader sitting next to a unit of measure, and any second-person direction to take, use, or adjust something. Reporting what a trial did reads differently from telling you what to do, and the check is anchored on that difference.
• Dose ladders: a low, normal, high table of amounts is a recommendation wearing a table's clothes. The gate greps for that shape.
• The disclaimer: present on the page, matching the locked wording exactly, on every page without exception.
• Structured data: the machine-readable block has to parse as valid JSON, and its author has to be the organization, never an invented person.
• Tag balance: every inline emphasis tag opened has to be closed. This check was added the day sixteen unclosed tags were found across ten pages, bleeding bold type through everything below them.
• Title and description limits: both authored per page and both inside their character ceilings, so a search engine is never left to invent a summary for us.
• Citation presence: a page carrying a sources section has to carry at least one resolving DOI, PubMed ID, or regulator link. The gate cannot tell you a citation is true, only that the page is not empty of them.
• Canonical and index state: one canonical address per page on the apex domain, and no leftover flag telling search engines to ignore a page we meant to publish.

Each of those lines has a story: almost every check on the list was written the day something got through, which is why the list grows and why none of it is aspirational. The script either passes a page or it does not.

How does a correction get made?

The page gets fixed, at the number: when a figure is wrong, we change that figure where it lives, in the body text and in the references, rather than bolting a note onto the end and leaving the error in place above it.

The change gets dated: the last-updated line on the page moves to the day of the fix. That line is the reader's handle on how fresh a page is, so it moves only when the content actually moved.

The correction gets recorded, not buried: a corrected figure is written into the correction log below, saying what was wrong, what it is now, and when it changed. Quietly editing a wrong number out of existence hides the one thing a reader most needs to know about a source of numbers, which is how often it is wrong and what happens when it is.

A corrected page re-runs the full check: it goes back through the scripted gate and the independent citation audit exactly like a new page. A fix that was never re-checked is just a second draft.

Claims that can go stale carry a re-verification date: anything time-sensitive, an approval status, a trial still running, a label that gets revised, is marked for re-verification ninety days out. Re-verifying means pulling the source again and reading it, never touching a date field to make a page look current.

How to report an error: we would rather hear it from you than leave it standing. There is no dedicated corrections address published yet and we are not going to invent one to fill this paragraph: the route is a reply to the weekly briefing, which reaches us directly. The list is not open for signups at this moment, and the form on this page says so honestly instead of pretending. When a direct address exists, it gets published in this section, and this section is the only place it will come from.

The correction log

Nothing logged yet: no correction has been issued since this standard was published on 27 July 2026. When one is, it appears here with the page it belongs to, what changed, and the date, and it stays here afterwards. An empty log is only meaningful if it is the same log that would show the entries.

What this standard does not cover

It checks pages, not the world: we verify that a figure matches the document it came from. Whether that document is right is the literature's argument, and where the literature contradicts itself we say so on the page rather than quietly pick a winner, which is the whole shape of the page on why two dose charts disagree.

It does not make a page permanent: labels get revised, trials report out, and a page that was accurate on the day it shipped can drift. The correction route above is the mechanism for that, and it depends partly on readers telling us.

It is not a promise about anyone's body: nothing on this site is personalized, because nothing on this site knows anything about you. Every page published under this standard is listed in the library, and every one of them is written to be read, checked, and argued with.

Last updated: 27 July 2026.