A duration score rejects an expected 88% of selected circular Kepler DR25 recoveries at b ≥ 0.8, log density scatter 0.3, and a 5% training budget

We test a duration and density score against supplied corrected fitted impact parameter in 25,546 selected Kepler DR25 pixel injection recoveries. All injected orbits are circular. The duration score uses an input density reference as extra information. Independent normal errors have specified standard deviation 0.3 in natural-log density. Five folds keep whole sky groups out of training. Each threshold has a 5% training rejection budget for true b below 0.8. The duration score rejects an expected 88.07% (87.13–88.95%) of cases with true b at least 0.8. Fitted b rejects 7.54% (6.30–8.65%). The intervals use 1000 paired sky-group bootstrap draws with threshold training repeated in every draw. The paired gain is 80.52 percentage points (78.92–82.32). Both held-out lower-class rejection rates are about 5%. The expected high-b share after the duration cut is 1.38%, compared with 9.77% for fitted b. Exact density gives 94.64% high-b rejection. We also report the response to log density scatter from zero to one, two secondary tests, and six subset checks. This is a conditional benchmark of an established geometry score. The stated error law does not describe measured stellar errors. The results do not establish performance for eccentric or real planet samples. Full source tables, code, decisions, group draws, figures, and numerical audits are included.

Kepler DR25 · transit geometry · impact parameter · stellar density · pixel injections · held-out validation · group bootstrap · selection