How we grade evidence

How we grade evidence

There are exactly three grades. They are deliberately few: a longer scale invites false precision, and every additional grade is one more judgement call to get wrong on every page that renders one. A grade is assigned to a specific claim, such as "supports gastrointestinal mucosal healing", not to a compound's reputation.

Human RCT evidence

At least one published randomized controlled trial in humans reports the outcome. Registrational trial programmes qualify. A trial that is registered but not yet published does not.

Animal studies only

The evidence is animal work in living animals, or laboratory work in cells, with no controlled human trial of the outcome.

Anecdotal reports only

Case reports, retrospective chart records, uncontrolled series, open-label reports without randomization, or community report only.

Two rules govern how the grade is picked:

• A compound's overall grade describes its typical claim, not its best claim. One small human record does not make a compound with an otherwise entirely preclinical file a human-trial compound.
• Grade down, never up, when uncertain. A weak human observational study does not lift a claim to "Human RCT evidence". Where the honest answer is that no human efficacy data exists, the page says exactly that and grades the claim accordingly. A claim with nothing behind it is written as unestablished rather than propped up with a tangential paper.

Why we also show study design

A three-value scale tells you how strong the evidence is and says nothing about what kind it is. Because there are only three grades, the weakest two absorb everything that is not a human trial. That is how a badge reading "Animal studies only" can end up over a citation list that is entirely cell-culture and review work. That is a true statement about strength and a misleading impression about substance.

Rather than add a fourth grade, every graded claim prints the study designs of the papers behind it, directly under the badge. "Anecdotal reports only · in vitro" and "Animal studies only · animal, review" say what the grade alone could not. The six designs are:

• human RCT
• human observational
• animal
• in vitro
• review
• product labelling

Every citation on the property declares its design at authoring time, and the loader refuses to publish a page whose citation carries a design outside that list. The designs you see are recorded, not inferred from the title.

The honesty rule

A claim graded "Animal studies only" has to cite at least one primary animal study. If everything behind a claim is cell-culture work or a review article, the claim is "Anecdotal reports only", not a weaker animal claim. Telling a reader the evidence is animal research when no animal experiment is cited overstates what exists, in the one direction that matters.

Where that rule is currently a build gate, and where it is not. On our goal, stack, and how-to pages the rule is enforced mechanically: a claim that breaks it fails the test suite and cannot ship. On the individual compound pages it is applied by review rather than by the build, and as of the review date above, 13 claims there carry "Animal studies only" over citation sets whose papers are review articles and cell-culture work summarizing animal research rather than primary animal experiments. Bringing the compound pages under the same automatic gate is open work, and it is tracked as such. (This count is not decorative: an automated check compares it against the library on every build, so if it drifts, this sentence fails before it can mislead you.)

This is exactly why the design line under every badge exists. Where a grade is coarser than the evidence beneath it, the designs printed under the badge show you directly. You do not have to take the grade's word for it.

How we cite

Every claim carries a PubMed identifier or a DOI. Nothing is ever cited from memory, and no citation is written to fill a gap. If a claim has no paper behind it, the page says the claim is unestablished instead of reaching for a tangential one.

Outbound links are not written by hand at all. An author records an identifier; the software builds the link from it. A page's prose can only cite by referencing one of those recorded identifiers, and before a page is allowed to publish, its body is rendered and every resulting link is checked against the set of links derived from its own identifiers. A hand-typed web address, a Markdown link, or a raw anchor tag in the body is a validation failure, not a style problem. That is what makes the promise structural rather than editorial: every claim links to its PubMed or DOI source, never a vendor and never a blog, because the system can produce no other kind of link.

How we verify citations

A well-formed identifier can still point at the wrong paper. So identifiers are not merely checked for shape. They are re-fetched. A verification pass takes every identifier on the property back to the registry that issued it, and we will be exact about how far that goes, because "verified" is a word this page has to earn.

For a PubMed identifier the pass pulls the live record and confirms it exists, then compares the title and the year the registry returns against what the page prints beside the citation. It also pulls the journal and prints it in the report for a person to read, but it does not compare that one automatically: our pages spell journal names out in full and PubMed returns them abbreviated, so a machine comparison would flag hundreds of perfectly good citations and teach us to ignore it. For a DOI, the pass confirms with the registry that the DOI is registered and resolves; DOI resolution hands back no metadata, so a DOI's title, journal, and year are checkable only by a person. Today every citation on this property is a PubMed identifier.

The title comparison is fuzzy and deliberately report-only. It flags a citation for a human to read rather than failing the run, because subtitles, diacritics, and errata-updated records produce false alarms and an automated verdict on those would be worse than no verdict at all. What the pass fails on is the thing it can be certain about: an identifier that resolves to nothing.

The pass runs against the live registries rather than in our automated test suite, on purpose: a network outage should read as "not yet checked", never as "this citation is fake". It reports those two outcomes as separate results, and it is re-run before any content change ships and at every scheduled review.

Why verification is not enough

Verifying an identifier proves the paper exists and that we quoted it correctly. It does not prove the underlying statement of fact is true. A real paper, faithfully quoted, can restate a regulatory or anti-doping fact incorrectly. If we take that characterization at face value, the error arrives on our page wearing a citation that checks out perfectly.

This is not a hypothetical. It happened here. Two statements of regulatory fact on this site passed a full identifier-verification pass. The papers were real and the quotations were accurate, and both statements were later found to be wrong about the regulatory fact itself. They were corrected by going to the primary document. A process that had declared itself clean had shipped two false statements, and the check that was supposed to catch them was structurally incapable of doing so.

The rule that came out of that: every regulatory, approval, enforcement, and anti-doping claim is verified against the primary document itself, and never against a review paper's characterization of one. The primary document is the World Anti-Doping Agency Prohibited List, the FDA communication, or the approved product labelling. Anti-doping section numbers are read out of the current Prohibited List, bullet by bullet, and recorded so that a wrong section fails our checks as loudly as a missing one. When the List is reissued, it is re-read rather than inferred.

We would rather tell you where this process has failed than imply it cannot. A methodology page that presented citation verification as sufficient would itself be a false claim about the process.

How often we review

Every content page carries a "last reviewed" date, printed near the top of the page. It means something specific: on that date, the claims, grades, and citations on that page were re-read against the current literature and the current regulatory position, and the citation verification pass was re-run. It is not the date the file was last edited, and a typo fix does not move it.

Two automated gates keep that date honest. A page whose review date is more than 12 months old fails the build, so a stale page cannot quietly stay published. A page dated in the future fails the build too, because a review date can only describe a review that has already happened.

A correction also moves the date. If we fix a factual error on a page, that page is re-reviewed and re-stamped, because the old date would otherwise claim we had checked the corrected text when we had not.

What we check automatically

The following are enforced by our automated test suite. Each one blocks a release rather than producing a warning, because a compliance rule that depends on an author remembering it is not a rule:

• No commercial pointers. No page, template, or helper on this property may name a seller or give any guidance about acquiring a compound, not even generic advice about how to evaluate one. The check scans every content file and every template.
• No preparation or administration instruction. No technique, no equipment guidance, no arithmetic that converts a concentration into a measured amount. Describing what a published trial administered is a fact about a study; walking a reader through doing it is out of scope here at every level of detail.
• No personalized dosing. Phrasings that address a specific reader's regimen fail the build. Published trial amounts are reported as what a trial did, never as what you should do.
• Anti-doping status, stated for every listed compound. Where a compound appears on the current Prohibited List, every page that discusses it names the section it falls under, and the section is checked against a table read from the List itself.
• No dangling cross-references. A page cannot link to a compound the library does not hold, so an internal link cannot rot into a dead end when content changes.
• A depth floor per page. Each page type has a minimum word count, section count, citation count, and number of graded claims. A thin page fails rather than publishes.
• Citation integrity. Every citation identifier is well formed and typed; every reference in the prose resolves to a declared citation; every declared citation is actually used; and no page may emit an outbound link that was not derived from one of its own identifiers.
• Review-date freshness. Nothing older than 12 months, nothing dated in the future.

What we deliberately do not do

These are not gaps we intend to fill later. They are the boundary of what this property is for:

• No personalized dosing. What is appropriate for you is a clinical judgement about you, and we know nothing about you.
• No commercial guidance. We name no seller and rank no product. We do not tell anyone where to obtain anything.
• No preparation or administration procedures. That belongs to a licensed provider and a pharmacist.
• No lab-result interpretation. Our Stack Builder never reads or parses a lab document you upload.
• No selling and no prescribing. Peptide Health Lab does not sell peptides and cannot write a prescription.
• No treatment relationship. Reading this site does not make anyone your clinician.

Common questions