What is checked, and what is still open.
Every number on this page was produced by the build that published it. None of it is typed. If a figure here is wrong, the build that wrote it was wrong, and a gate exists to say so.
Most sites in this field publish what they have finished. This page also publishes what they have not. The open items below are real, they are counted, and they are printed on every deploy so the number cannot drift quietly while nobody is working on it.
One limit stated plainly, because it is the most important thing on the page: an automated check can confirm that a citation exists, that it is not retracted, that it is attached to the claim it belongs to, and that a grade does not exceed what its endpoint The outcome a trial measures to decide whether the treatment worked, such as heart attacks, blood pressure or a symptom score. allows. It cannot confirm that a paper actually supports the sentence it sits under. That judgment is adversarial and it is recorded per record. Most of it was made by a language model reading against the papers rather than by a clinician, every record says which, and the split is one of the figures below.
Every letter the database serves, counting a per-indication grade as its own letter. A record graded once at the top and three times by use contributes four.
Each letter records whether it was earned on a hard clinical outcome, an objectively measured function, a surrogate biomarker A measurable sign in the body, such as a blood test value, used to track a condition or the effect of a treatment. or a self-reported scale. A surrogate or a self-report cannot reach Grade A, and the build refuses to publish a letter that exceeds what its endpoint allows.
D covers two different findings: no supporting human trial exists, or one was run and found no benefit. Both are published rather than omitted.
The supplement table alone. One at the top grade out of 109 is the honest shape of this evidence base.
Every one re-checked against PubMed on every deploy for retraction and for expressions of concern. A grade resting on a withdrawn paper fails the build.
The publish pipeline. A failed check stops the deploy: nothing reaches this site without passing all of them.
An automated check cannot tell whether a paper actually supports the sentence above it. That reading is adversarial and it is recorded per record: what the trial pre-specified, whether it met it, whether the claim rests on a primary or a secondary result, and whether the letter is the honest ceiling. Every record names who did the reading, and the split is published below rather than left to an assumption. Most of it was done by a language model reading against the papers, not by a clinician.
The tier where a judgment is mandatory rather than optional. Editing the evidence fields of one of these expires its judgment and the build fails until it is judged again.
What the trial actually administered: form, dose, population, duration. The ones naming a gap say plainly that the product on a shelf is not the thing that was tested.
A letter cannot move silently. Every change is dated and published, including the ones that lowered a grade after a closer reading.
These are counted, not estimated, and they are printed on every deploy.
Every graded record on this site has been read adversarially against its own papers, at every tier. Coverage is not the open question here. Who did the reading is: see the line below.
This is the real open item, and it is the one an outside reader should weigh most. A judgment marked as editorial review was made by a person. The rest were made by a language model reading adversarially against the papers, are labeled that way on the record itself, and have not been reviewed by a clinician. Machine coverage is complete; independent clinical review is not.
Long-standing clinical observations labeled as such rather than dressed in a reference that does not support them. Honest, and still unfinished.
Each states that it rests on pharmacology rather than a study, with the absence recorded. Stating the basis is the floor here; it is not the same as having the trial.
Nothing is served that is not on this list. Anything else in the repository is deleted before the site is built.
Generated by render_integrity_index.py and verified by gate_g66_integrity_index.py, which re-derives every figure above from the same files and fails the deploy if any one of them disagrees. No number on this page can be typed by hand.
It is not a quality measure. A longer reference list is not better evidence, and a page can cite forty papers and still rest on none of them. The grade on any entry here comes from what its trials actually tested and in whom, never from how many were attached to it. Counting references measures breadth, and breadth is the least interesting property evidence has.
It is not fixed. The figure moves whenever a paper is added, retired or superseded, which is why every statement of it here carries the date it was generated. Two documents written a week apart may honestly carry different numbers, and saying so is more accurate than forcing them to agree.
It is not a claim that every paper was read the same way. What each citation was checked for, and what remains unverified, is set out separately in the citation audit, which lists every defect found in this site's own citations rather than only the ones found in other people's.