This preprint presents an independent research project on long-context large language model (LLM) auditing. The study introduces a Logical Needle-in-a-Haystack (L-NIHS) stress test to examine whether long-context LLMs can preserve the correct reasoning path from buried evidence to a final numerical closure under different prompting conditions. Instead of evaluating only whether a model can retrieve relevant information, the project focuses on whether the model can maintain the correct closure decision when persona framing and late-stage distractors are introduced. The experiments use two synthetic regulatory-style corpora and a Reader–Judge evaluation pipeline to analyze closure substitution, evidence selection, output length, and downstream rubric-based verification behavior across multiple solver models and prompting conditions. This work is intended as an independent research preprint and has not been peer reviewed.
Junzhe Cai (Tue,) studied this question.