A 1.8 Earth-radius crossing term did not reduce test error in a fixed model for 347 selected Kepler pairs

We tested whether a binary radius term reduced orbital spacing prediction error for 347 observed adjacent Kepler pairs at 264 stars. A pair crossed the planned 1.8 Earth-radius cut when its planet radii were on opposite sides. We joined a saved NASA DR25 snapshot to Berger et al. (2020) radii. We set adjacency before endpoint cuts. The baseline model used continuous planet sizes, inner period, and observed planet count. Each star had equal weight. Ten test folds kept all pairs at a star together. The full-to-baseline test root mean square error ratio was 1.00160, or 0.16% larger error. A 2,000-draw star bootstrap that refitted all training models gave a 95% range of 0.99171 to 1.01166. The adjusted crossing coefficient was -0.113 in log spacing, with a star-cluster 95% interval of -0.281 to +0.055. At a 1.6 Earth-radius cut, the coefficient was +0.262 [+0.070, +0.454]. Its test error ratio was 0.98704, but a conditional bootstrap range included one. Thus the fitted contrast depends on the cut. An exact example also shows that a period-dependent radius label can select wider pairs when all radii are equal. The package preserves inputs, source hashes, code, all planned checks, test predictions, figures, and a second parser. This is a test of an added term in a fixed model and selected observed sample. It does not estimate a causal effect or an intrinsic planet population.

Kepler · radius valley · orbital spacing · cross-validation · planetary systems · selection effects · reproducible research