The evidence ladder
Every study carries a single number from 1 to 5. It says how much the evidence weighs — not how good the news is.
| 5 Repeated evidence | Several studies in people combined into one analysis. What comes out of it is not the result of a single research group or a single lucky finding.
Closing line on the study page: “What this means for you: this is established.” |
| 4 Large controlled trial | One controlled trial in people with a hundred participants or more: there was a comparison group, and the group was large enough to mean something.
Closing line on the study page: “What this means for you: this demonstrably works.” |
| 3 Small study in people | Research in people, but small or without a comparison group. Enough to suspect something, too little to promise anything.
Closing line on the study page: “What this means for you: first signals, nothing more.” |
| 2 Animal study | Studied in animals. However attractive the result: most substances that work in mice do nothing in people.
Closing line on the study page: “What this means for you: promising, but mice are not people.” |
| 1 Laboratory | Studied on cells in a dish. Everything starts here, and most of it gets no further.
Closing line on the study page: “What this means for you: nothing yet — this is laboratory work.” |
Who assigns the number
Not a language model, but a fixed rule. PubMed records for every article what type of study it is and whether it was done in people, animals or cells. That judgement comes from the National Library of Medicine, not from ProofDigest; we translate it into the ladder above.
One rule comes before all others: who it was studied in weighs more than how it was studied. A meta-analysis of animal experiments is a meta-analysis — and it is about mice. Here it gets level 2, not level 5. That distinction is exactly where this market makes its money.
What is not in here
Review articles carry no evidence of their own; they retell, and they belong in the knowledge base. Research in children, pregnant women, athletes or patient groups unrelated to the question is left out. And where the classifier is not certain, a study gets no number at all rather than a guessed one.