A large effect in a weak study and a small effect in a strong one are different findings. Averaging them into a single score destroys the distinction, so this reference refuses to. Every indication carries an evidence grade for how the claim was studied, and an effect signal for what was observed.
The grade is about study design
A is a randomised or controlled human trial. B is preliminary human data — pilot, open-label, case series, observational. C is mammalian in-vivo work. D is mechanistic: in vitro, cell culture, or inference from a pathway. A narrative review never lifts a grade on its own, because a secondary summary is not primary evidence.
The effect signal is about magnitude
Large, Moderate and Small describe what the cited work measured. Mixed records that studies disagree, and No effect records that something was studied and nothing was found. Those last two exist deliberately — an archive that cannot record a negative finding is not an archive.
An effect signal never renders without its grade. Large on its own is a lie by omission; D · Large reads correctly as a big effect in cell culture.
What is derived and what is not
Grades on this build are currently derived rather than authored: each citation is classified by study design from its metadata, and a compound's badge takes the strongest design among its references. The interface says so wherever it appears. The published three-axis rubric is designed but not yet implemented, and until it is, no entry should be read as rubric-graded.
The full rubric, the effect scale and what is still outstanding are set out on the methodology page.