LSP-Orbital v4.3b -- revised with honest satellite-level validation. This paper presents a lightweight geometric corrector that learns the residual structure between Keplerian propagation and SGP4 using 29-dimensional Lo-Shu polynomial features. Key design: train/test sets use strictly disjoint satellites (80/20 satellite-level split, no data leakage). Results on 50 held-out satellites (296 LEO samples): 64.1% mean error reduction, 91% win rate, 9% failure rate (explicitly reported). Time-stratified training (SHORT/MID/LONG layers) increases tangential R2 from 0.33 to 0.668. Empirical finding (not proved theorem): Lo-Shu row-imbalance predicts SGP4-Keplerian residuals across all LEO families. Limitations explicitly stated: SGP4-residual scope, no BSTAR, no future-TLE validation. Supersedes v4.1 (10.5281/zenodo.20080799).
Yao-Kai Kao (Fri,) studied this question.