Narrative Context, Direct Behavioral Scripts, and Downstream Shopping Choices in Large Language Models: A Preregistered Two-Study Audit

This record contains the complete preprint and analysis reproducibility package for “Narrative Context, Direct Behavioral Scripts, and Downstream Shopping Choices in Large Language Models: A Preregistered Two-Study Audit.” Reports that emotionally evocative narratives alter later large-language-model choices leave at least three explanations unresolved: a state-like contextual effect, semantic content, or an executable behavioral script embedded in the prompt. We conducted two preregistered studies of shopping choices. Study 1 was a high-fidelity procedural bridge using DeepSeek V4 Flash in 147 fresh sessions. Study 2 used 2,120 fresh sessions split equally between GLM-5.2 and DeepSeek V4 Flash to separate matched neutral context, emotion labels, objective adverse-event content, rich narrative, abstract decision-bias cues, and direct shopping instructions. Rich narrative did not support the predicted reduction in terminal-basket health relative to matched neutral context, whereas direct shopping instruction produced a large, cross-model and measurement-robust decrease. No preregistered rich-narrative persistence signature was established. Under the tested models, prompts, catalog, and measurement rules, the evidence favors sensitivity to explicit task-specific instructions over a robust narrative-induced deterioration in shopping composition. These findings do not establish the absence of internal model state or subjective emotion and do not invalidate prior work. Files included: paper2_preprint_v0.2.2.pdf: complete 15-page preprint. paper2_public_release_v0.2.3.zip: analysis-ready data, code, figures, protocols, claim-evidence documentation, integrity manifests, and reproduction instructions. The archive is an analysis reproducibility package rather than a complete collection of raw vendor API logs. API credentials, account information, private configuration, and unrelated pilot or superseded execution materials are excluded.