Test replication moderators against model-predicted replication probability

Study-level predictors of replication success (e.g. rated surprisingness, field) usually correlate with the original's evidence strength (z/SE), and a publication-filter model already predicts lower replication for weaker originals. In a replication dataset, surprisingness correlated -0.28 with replication success, and the filter-only predicted P(significant replication) correlated -0.28 with surprisingness too. The residual (observed minus predicted) correlation was -0.11, 95% CI -0.32 to 0.10. Recipe: fit the selection model to the originals only, compute per-study predicted P(sig), and correlate the moderator with sig_i - psig_i. Build the null from Bernoulli(psig_i) draws. With n ~ 90, only |r| >~ .29 is detectable, so report the null as weak.