Replaying Twenty Questions with Nature: How Much Information Survives Replication in Social Science?

The replication crisis has revealed systematic limits in how social-science findings transmit information about underlying effects. Yet current debates and diagnostics do not cleanly localize where along the evidential chain information is lost as effects move from data to publication and onward to downstream evaluation. We introduce a novel information-theoretic framework that models the research pipeline as an information channel and quantifies how much variability in a corpus-defined empirical signal proxy is (i) captured at publication, (ii) retained under direct replication, and (iii) transmitted end-to-end through the observed evidence pipeline. The framework defines three normalized indices—Capture, Retention, and downstream information-throughput—computed from mutual information among a signal proxy, published evidence, and replication outcome. Applying the method to four major replication corpora spanning psychology and experimental economics reveals distinct information-flow profiles across research settings. Publication outputs capture only a modest fraction of signal-proxy variability, and the amount of information carried forward into replication outcomes differs substantially across corpora. These results demonstrate that replication rate alone does not characterize how scientific information is preserved. The framework provides a new quantitative diagnostic for localizing information loss, comparing research pipelines on a common scale, and identifying where methodological reforms may have the greatest impact.