A geometric method for detecting LLM-generated scientific papers. We measure the directional reversal rate (F4) of sentence-embedding trajectories — how often consecutive text windows reverse direction in embedding space. F4 alone achieves AUC 0.921 across four LLM families, is temperature-invariant, and resists three rounds of automated self-revision. The paper also documents a length confound that inflates performance of section-level detection methods (AVG: F1=0.982 → balanced accuracy 60% after length correction), and recommends length-residualized evaluation as a reporting standard. Code: https://github.com/Faltland/trajectory-geometry-detector
Hans Joachim Koerber (Sat,) studied this question.