A circular eclipse model with 1.73% source light lowers prediction cost in two of four Kepler-1658 quarters

We tested the eclipse-shape evidence for dilution in Kepler-1658 (KIC 3861595). We used all ten public Kepler DR25 short-cadence files. Time-coverage cuts kept 66 complete eclipses and 74,960 flux samples in Quarters 2, 4, 7, and 8. We fit local quadratic baselines, common timing, circular eclipse geometry, and physical quadratic limb coefficients. The pooled fit gives an eclipsed-source light fraction of 0.01746. It lowers the archive-error fit cost by 20.50 relative to a model with light fraction 1. This is close to the published pooled result. The best fit reaches a central eclipse and zero intensity at the stellar edge. We then omitted each quarter, fit the other quarters, and predicted the omitted flux. Only local baseline terms were fit to test data. The fixed 0.0173 light model has lower prediction cost in two of four quarters. Its total gain is 9.22; a free-light model gives 9.24. Block score ranges include zero. Ten-minute residual averages have 1.65 to 1.99 times the scatter expected from archive errors. These tests reproduce the pooled fit gain but show limited prediction gain across observation blocks. They do not reject dilution or test the separate spectral evidence for a stellar binary. The package contains native data, source hashes, code, all block scores, and numerical checks.

Kepler-1658 · KIC 3861595 · eclipsing binary · dilution · short-cadence photometry · quarter prediction · correlated noise