Methods
The product is an epistemic filter, not a supplement shop. An edge is allowed onto the atlas only when we can name its design class.
The four stamps
Supported by an experiment or a tightly identified design. Still read the caveats.
Mendelian randomization. Good against confounding; not magic against pleiotropy.
Each side moves the other. A single arrow is the wrong picture.
The association does not survive the backdoor. Usually SES, outdoor time, or healthy-user bias.
What we will not do
- Promote a bivariate observational correlation as a treatment effect.
- Hide an interaction (T × cortisol) inside a slogan.
- Cite a paper we have not opened. Camerer did not run the testosterone ultimatum RCT; Eisenegger, Fehr and colleagues did.
- Pretend Inquire is full do-calculus. It is a backdoor scan on a curated graph.
- Give medical, legal, or financial advice.
How a new edge gets in
Extract X, Y, effect size, p-value if reported, design class, and N. Attach the DOI or PMID. Name confounders. Prefer RCTs and Mendelian randomization when they exist; keep observational work when it is the literature, and stamp it honestly. Future ingestion from PubMed and Semantic Scholar will write into the same contracts. Until then the seed inpackages/contracts/src/data/seed.tsis the source of truth. The public bibliography is /sources. The screenshotable grade table is /matrix.