How we grade evidence: four questions, never one score

The four grades together have a shape, and the shape is the information. Two compounds with the same “average” can be in completely different situations:

A star rating would put both somewhere in the middle. A reader deciding whether to trust an effect, or how cautious to be about harm, needs to know which situation applies. So the four grades sit side by side and are never added, averaged or ranked.

The grades describe the published evidence. They are not a verdict on whether a compound works, and not a judgement of the people who did the research: for BPC-157, see the Croatian lab.

Human clinical evidence

We take the highest level of human evidence that is both completed and published. A trial that is registered, or reported only as a conference abstract, does not count, because nobody outside can check its methods and results. BPC-157 is the clearest case: a randomised phase 2 trial in 53 people with ulcerative colitis exists only as a 2005 conference abstract, so the grade is “case reports”, not “phase 2”.³ “Authorised” means a medicines regulator has granted marketing authorisation somewhere for some use. It describes the whole programme of trials behind that decision, not every use people put the medicine to.

Preclinical volume

We count primary laboratory and animal publications: studies that report their own experiments. Reviews, commentaries and letters are left out, because a review that repeats a study is not a second experiment. The thresholds are fewer than 10, 10 to 50, and more than 50. For BPC-157 we searched PubMed, removed records tagged as reviews, comments, letters, editorials or errata, sorted the rest by hand, and counted about 160 primary laboratory and animal studies.¹

Source independence

We look at who wrote the primary studies. If more than about 70% come from one group (counting its collaborators), the grade is “single group dominant”. If two to four groups account for most of the work, it is “few groups”. “Broadly replicated” needs many independent groups and repeats of the key findings. Independence matters because one group’s shared methods, models and expectations can carry the same error into every study it publishes; a second laboratory with its own methods is less likely to repeat that error. When independent scientists have tried to repeat published preclinical findings, many did not hold up.^4,5

Safety characterisation

We use human data only. Animal toxicology tells you what to look for, not what happens in people. “Uncharacterised” means no systematic human safety data. “Partial” means some exposure in people, but short-term or small. “Well characterised” needs large trials plus monitoring after the medicine reaches the market, because only then are enough people exposed for long enough to see uncommon harms.

No rubric fits every case. Two examples from this site:

• Retatrutide has published phase 3 results, but every human trial was sponsored and largely written by one company, while about 15 independent laboratories have each published one or two small animal or tissue studies.⁶ “Few groups” is the closest fit for independence, but it is a judgement: the key human findings have not been repeated outside the company. The compound page says so in its reason for the grade.
• Semaglutide is graded “broadly replicated” although the manufacturer paid for every pivotal trial. The compound page gives the reason: the core effects on blood sugar and weight have been reproduced in studies without company authors, and the main safety questions were tested by independent academic groups and by regulators.^7,8,9 Funding and independence are related, but they are not the same question (see conflicts of interest).

In such cases the written reason on the compound page carries more information than the grade itself. See the retatrutide and semaglutide pages.

A label goes on every claim about what a compound does or how it works. It describes the claim, not the compound: on a single page, one claim can be marked “Animals” and the next “Human trials”. A claim that reports a rat study is labelled “Animals” even when the compound is an authorised medicine, because the label is about that sentence’s evidence.

The mechanism chain on a compound page is read link by link. A chain can be well shown at the first step (the molecule binds its receptor, measured in cells) and only hypothesised at a later one (that this binding is what heals a tendon). The weakest link limits the whole chain: a pathway is only as well established as its least-tested step. That is why each step carries its own label.

Counting is not quality

“Extensive” preclinical volume means there are more than 50 primary studies, not that they are good. In animal research many studies do not report randomisation, blinding or how their size was chosen, and published animal studies lean towards positive results.^10,14 When independent scientists tried to repeat landmark preclinical findings, many did not hold up.^4,5 That is why preclinical volume is always read together with independence, and why the translation gap explains what animal studies can and cannot show.

Known is not complete

“Well characterised” safety still leaves rare harms to be found. Semaglutide was authorised in the EU in 2018 after large trials, and in June 2025 the EMA’s safety committee concluded that NAION, an eye condition that can cause sudden loss of vision, is a very rare side effect: one that may affect up to 1 in 10,000 people taking it.^15,11 Trials of a few thousand people can easily miss a harm that rare; wider use over years can find it.

This rubric is not GRADE

Guideline groups worldwide use GRADE to rate how much confidence a body of evidence deserves for one specific question, from “high” to “very low”. It starts randomised trials high and observational studies low, and moves them down for study limitations, inconsistency, indirectness, imprecision and reporting bias.¹⁶ Our rubric does something coarser and different: it describes the overall state of the evidence on a compound, across all its claimed uses. It does not tell anyone whether to use a treatment, and it cannot replace a GRADE assessment for a clinical question.

Grades move in both directions, and we would report either.

• Human clinical evidence moves up when a trial is published in full. If the old BPC-157 phase 2 trial were published with its full data, it would count; so would the results of a registered trial, once published.³ It never moves up for a press release or a registration.
• Independence moves up when other laboratories repeat the key findings, and stays where it is when new papers only review old ones. Between 2024 and 2026, most new BPC-157 records in PubMed (34 of 52) did not list the Zagreb group’s leader as an author, but more than half of those were tagged as reviews, comments, letters or editorials, which do not count as repeats.¹
• Safety moves from “uncharacterised” to “partial” only with systematic safety data from people, collected in a study, not with an absence of reports.
• Any grade moves down if we find we counted wrongly, or if a published study is retracted.

The Croatian lab sets out the five things that would change the BPC-157 picture.