Replication package for "Differential Effects of LLM Persona Elements on Do-main-Specific Decision-Making: Estimating an Element–Domain Elasticity Matrix"

Replication package for Seo, Ko, Moon, Lee and Kim, "Differential Effects of LLM Persona Elements on Domain-Specific Decision-Making: Estimating an Element–Domain Elasticity Matrix" (submitted to Systems, MDPI, 2026). The study compares how strongly six persona elements (demographics, Big Five traits, basic values, behavioural history, narrative identity and social identity) are associated with consumption, political and information decisions in a human baseline (N = 364) and in persona-conditioned large language models. Contents: - data_llm/: all LLM responses, prompt plans and gzipped call logs for the Persona-1M panels (Llama-3.3-70B-Instruct, Qwen2.5-72B-Instruct; 996 personas each), the no-persona control, the reasoning-model arms (DeepSeek-R1-Distill-Llama-70B; Qwen3-32B with thinking on and off) and the prompt-length ablation (full vs. 47 matched attributes). - plans/: hash-frozen analysis plans, configurations and SHA-256 freeze logs, including the human analysis plan SAP v1.4 (author e-mail redacted; also registered on OSF). - code/: analysis and runner code (Python, R). - results_aggregate/: aggregate results behind the tables and figures (cell matrices, pair-level correlations, bootstrap replicates of correlations and cells, sensitivity analyses). Not included: human participant data and digital-twin data, which contain participants' own responses and life narratives. These are available from the corresponding author on reasonable request, in line with the conditions of participant consent. See README.md for the folder structure, the correspondence with the article and the full list of exclusions; MANIFEST.csv lists the size, SHA-256 hash and source of every file. Data are licensed CC BY 4.0 and code MIT. Persona attributes derive from the MatrAIx Persona-1M dataset and remain subject to its licence.