The Bias Tax in Long-Context LLM Auditing: Closure Substitution and Verification Overhead under Persona Prompting

This preprint investigates how persona prompting affects reasoning reliability in long-context LLM auditing. We introduce a Logical Needle-in-a-Haystack (L-NIHS) evaluation setup in which models must recover distributed evidence, reject a late-stage distractor, and preserve the correct reasoning path to a numerical closure. Across two independently constructed long-context corpora, four solver families, six prompting conditions, and 2,095 main trials, with an additional 566 inter-judge rescoring trials, we examine how prompt framing influences closure accuracy, evidence use, output characteristics, and downstream verification burden. The results suggest that persona conditioning can distort the final commitment even when relevant evidence has been successfully retrieved, and that such effects vary substantially across models, prompting strategies, and corpus structure. This version is a revised and condensed preprint with updated cross-model experiments, inter-judge validation, and statistical analysis. Supplementary materials, including corpora, prompts, experiment logs, and analysis resources, are linked in the manuscript.