Methodology

Three counts, read out of the published compound pages rather than typed here — which is the first thing this page is demonstrating.

Every reference on this site is resolved through the database that owns it, in this order. A citation that cannot complete these steps does not ship — it is not softened into “research suggests” and it is not carried over from another article.

1. Complete

   Search the record, do not recall it

   The starting point is a query against the registry or index, never a paper someone remembers. An identifier written from memory is the single most common way a fabricated citation enters a page.

2. Complete

   Pull the record's own summary

   Authors, year, journal, title and identifier come back from the database itself and are stored exactly as returned. A DOI is read out of that record; it is never assembled from a pattern.

3. Complete

   Open the abstract, not the summary

   The number the page is about to print has to appear in the source. “About 15%” attached to a paper reporting 14.9% is a rewrite, not a citation.

4. Complete

   Check the claim, not the topic

   A source that is merely adjacent to a sentence does not support it. The question is whether the claim follows from the record, and a source that only shares a subject is a wrong-source failure.

5. 5 — Current step

   Re-resolve before every build

   The stored reference is checked back against the live record automatically, so a citation that has drifted from its source fails the build rather than sitting on the page.

Editorial discipline is not a control. The check that actually keeps a fabricated reference off this site is automated and blocking, and it runs in two layers because the first layer alone is not enough.

Two labels, kept apart on purpose — what the evidence supports, and what you can lawfully obtain. The editorial policy states what each label means. What follows is the whole register with both labels attached, so the rubric can be audited rather than believed.

Ordered by how far the published human evidence goes, strongest first — then alphabetically. Nothing else moves a row. The sort reads the evidence tier and the compound's name. There is no partner flag, priority number or featured slot in our content model for it to read, and none will be added — an evidence-graded list a commercial relationship could reach would make the whole rubric a decoration.

The two labels are independent, and the pairs that fall out of that are the reason they are not collapsed into one score:

• Real evidence, still out of reach — Survodutide, Cagrilintide and Retatrutide. Trials have run in people; a US reader cannot lawfully obtain them today. A single score would read these as mediocre.
• On sale, never tested in people — BPC-157 and TB-500. Widely marketed, and the published work is animal or cell data. A single score would read these as merely unproven rather than untested.

A grade is a statement about the published record, so only the published record moves it. The four steps and what each one requires:

The most useful sentence on this subject is usually about something that does not exist, and it is also the easiest one to overstate. Every absence claim here is written to a fixed shape — what was searched, where, and when — and the editorial policy sets out why.

The procedural half of that, which is what this page owes you:

• The search is run against the database that would hold the thing, not against a general web search. “Nothing on PubMed” and “nothing on the internet” are different claims and only one of them is checkable.
• The date is the date the search ran, and it is printed. An absence claim decays; a dated one decays visibly.
• An absence of reported harm is recorded as an absence of looking whenever that is what it is. Nobody having studied an interaction is not evidence that the interaction is safe.
• A negative result is published as a result. A search that finds nothing is a finding, and dropping the topic instead would leave the reader to assume the opposite.

A site about compounds most people cannot lawfully buy is defined as much by what it refuses to print as by what it grades.

A dosing tool is the highest-liability surface on this site, because it does not describe a number — it hands one to a reader. The rule, stated in full on the editorial policy, is that our tools convert and never recommend. Mechanically, that means:

You supply the dose

Every field ships empty. There are no presets, no “typical” placeholder, and no default that a reader could mistake for a suggestion.

We supply the arithmetic

The output is your figure converted into concentration and syringe units, with warnings when it exceeds the vial or the syringe. If a tool cannot compute without inventing a number, it computes nothing.

A date on a page is a claim about work, and a freshness stamp with no re-reading behind it is a lie with a date on it. What actually gets re-opened, and when:

It will. A number is read out of the wrong arm of a trial, a grade lags a readout, an availability label is a quarter stale. When that happens the fix is published as a fix — what was wrong, what it is now, and when it changed — on the corrections page, which is also where to send one. A claim on this site that you cannot trace back to its source is worth reporting, and it is the report we most want to receive.