How we decide what to publish

The Crossref gate

No paper reaches a published state until its DOI has been verified against Crossref — automatically, not by eye. We confirm that the DOI we hold resolves to a record whose title, first author, and year match the entry in front of you. On a pass, the entry earns a ✓ Crossref-verified mark and the raw Crossref record is stored for audit. On a failure, the entry stays unverified and the reason is logged. The gate never passes silently, and we never force it — an entry that cannot pass simply waits.

Primary-source verification

For the numbers that matter most — the ones people plan staffing around — a DOI match is not enough. Someone reads the actual paper. When a source sits behind a paywall or exists only in print, that has meant ordering it through interlibrary loan and reading the full text before a figure is allowed to anchor a model or a claim. Figures verified this way carry a PRIMARY-VERIFIED mark. Where we could not obtain the full text, we say so rather than paraphrase from an abstract or from memory.

Draft until reviewed

Content moves through three states. DRAFT means it has been assembled and its sources checked, but a physician has not yet signed off — it is marked with an unmissable draft strip and is not for operational use. Going from draft to published is a deliberate act gated on reviewed_by = BA plus a verified source; it can never happen by quietly flipping a status, and since migration 025 that gate is enforced by the database itself, not merely by the code path that normally performs it. Syntheses, cases, and the staffing framework all obey the same gate.

The third state is the honest one. Part of this library's editorial layer — the plain-language summary and the “so what” — was drafted by an AI model reading the published abstract, and we are ahead of our own capacity to read every one of those individually. Those entries are published, because a checked citation with a drafted summary is more useful to you than nothing at all, but they carry an AI-DRAFTED — PHYSICIAN REVIEW IN PROGRESS notice on the entry and in the library list, and it stays there until a physician has read that specific entry. The notice does not disappear when we publish; that is the whole point of it. Where it is present, treat the linked source as authoritative and the summary as a signpost. The number of entries still awaiting that read is tracked in the database and falls as the reading gets done — it is not hidden, and it is not the same claim as a verified citation.

We link, we don't re-host

We do not mirror or republish article text. Every entry points out to the publisher's DOI, and to PubMed or PMC where a free version exists — each link labeled free or may be paywalled so you know before you click. The library is an index and an honest reading of the evidence, not a copy of it.

The provenance chips, defined

• ✓ Crossref-verified — the DOI resolves to a matching Crossref record.
• PRIMARY-VERIFIED — a human read the full text and confirmed the figure.
• DRAFT — assembled and source-checked, awaiting physician sign-off; not for operational use.
• BA-CONSTANT — a placeholder number with no published basis, shown in the open and editable in admin; any calculation that leans on one is flagged low-confidence.

What counts as a legal requirement

Our regulatory survey reports how many of the 51 US jurisdictions carry an event-medical requirement. That number depends entirely on what counts as a requirement, so the rule is stated here rather than left implicit.

A jurisdiction enters the count only if the requirement appears in statute or in a promulgated rule. Agency guidance, planning guides, permit conditions and health-officer practice do not enter the count, however firmly they are applied in practice.

The District of Columbia is the case that shows why the line is drawn narrowly. DC requires every special event to have a health, medical and safety plan approved by its Department of Health, and its planning matrix requires a physician-staffed aid station at attendance thresholds lower than most states legislate. By any operational measure it is one of the most demanding jurisdictions in the country. It is nonetheless outside the count, because that requirement lives in agency guidance enforced at permit approval, and because the District's special-events licensing statute was repealed in October 2025.

Admitting DC on the strength of enforced guidance would mean applying a standard to DC that we have not applied to the other fifty. We surveyed the states for statutes and promulgated rules; we did not systematically survey all fifty for EMS-office guidance, county permit conditions or health-officer practice. Counting DC in without that work would inflate one jurisdiction simply because it is the one we happened to read most closely — a selection artifact, not a finding.

The consequence is that our headline number understates what is required in practice across the country. We would rather understate against a rule a reader can check than overstate against a judgment they have to trust. Where a jurisdiction imposes real requirements outside the count, we describe them in full and say why they sit outside it.

Where checking our work will mislead you

If you go to verify one of our citations, you may hit the same walls we did. These are the ones we know about, published so that a failed check does not read as a fabricated citation.

Connecticut §19-13-B87 does not resolve through the obvious search. The Connecticut eRegulations portal's Quick Search — Exact field returns "Unable to find section 19-13-B87" for that exact string. The section exists and is served normally through Content Search and through the Browse tree. Anyone verifying our most prominent regulatory claim — one physician per 1,000 attendees, the hardest ratio in the country — through the portal's front door will conclude the section is not there. Working route: Search → Regulations of CT State Agencies → Advanced Search → Content Search "mass gathering" → ten results covering §§ 19-13-B81 through B96.

A DOI can resolve and still not be the paper. We held a record for 10.1017/s1049023x0003661x, catalogued as a peer-reviewed paper. The DOI is genuine and resolves cleanly — to a congress abstract in a supplement, not to the paper. The full peer-reviewed study of the same work exists under a different DOI, with two additional authors and complete methods. A resolving identifier establishes that a document exists, not that it is the document you think.

Machine-extracted citations are where our worst errors live. Our District of Columbia row was populated by machine extraction from a PDF and flagged pending. It cited D.C. Code § 1-325.81 as the authority for a medical requirement. That section establishes a fire-department fee fund and imposes nothing on event organisers. The extraction had attached a plausible statute number to a requirement that statute does not contain, and it survived into a published table until the row was read by a person. The same review found the row understated what DC actually requires. A pending flag is not a placeholder — it is a claim that has not yet been checked.

Tables can encode meaning in glyphs that text extraction destroys. The District's medical asset matrix distinguishes required from recommended using two different marks — a letter in one font, a symbol in another. Copy the table as text and the distinction disappears entirely; every cell looks identical. We recovered it by reading glyph identity rather than the rendered character. Where we report a requirement as mandatory, that distinction has been checked.

What the checking has caught in us

A verification process is only credible if it produces findings against the people running it. Ours has. The corrections log records each one with attribution and severity; the pattern is summarised here because the pattern is the useful part.

We have published a citation that did not exist, assembled from the title of one paper and the authorship of another — it carried no DOI, which is precisely why nothing caught it. We have published a count that used the wrong unit, stating a number of states against a denominator of jurisdictions. We have published a figure whose encoding inverted its meaning, making the weakest standard in the country appear to be the strongest. We have cited a funding statute as the authority for a clinical requirement.

The common thread is derived data. Every one of those errors came from something computed, extracted or summarised rather than read. None came from a source a person had opened and reconciled. That is why our verification marks distinguish registry and primary-source checking from everything else, and why unresolved entries are counted in public rather than quietly omitted.

Corrections

Every page links its sources. If we got something wrong, the evidence wins — tell us and we fix it, and we log it publicly with attribution. The full corpus, including verification status for every row, is downloadable in machine-readable form from the corpus page.