The reusable core of Lupine is not a result, it is a discipline. Three techniques do\nmost of the work, and the same techniques are what kill our own bad claims.
\nElements (and potential families) differ in how many potentials exist for them.\nSample size alone tightens correlations. So any cross-group correlation is\nconfounded by n until proven otherwise.
\nThe control: down-sample every group to a common n (bootstrap), then compare. A claim\nonly survives if the effect persists at matched n.
\nA small fraction of corrupt records can manufacture a dramatic effect. Before\nbelieving any strong result in a noisy corpus, audit the records driving it.
\nCorrelations pooled across heterogeneous groups can invert when a confounder\n(element identity) is ignored — Simpson's paradox / Robinson's ecological fallacy.\nCorrect analysis requires stratified evaluation and random-effects meta-analysis\n(DerSimonian–Laird), not naive pooling. The formal spec went further and proved that\nthe specific Simpson's-paradox claim the original paper made\ncannot arise under the real causal graph.
\nStrong results in a noisy corpus deserve suspicion before celebration. Every\nrefuted claim is published with its confounder named (the\nconjecture ledger is the register). A method that catches its\nown mistakes is the asset; the surviving claims inherit its credibility.
\nSloppy models: Brown & Sethna (2003), Transtrum & Sethna (2011). Ecological fallacy:\nRobinson (1950), Bickel et al. (1975). Meta-analysis: DerSimonian & Laird (1986).\nCausal inference: Pearl (2014).
\n"}