Analysis-ready data for: A hybrid Monte Carlo–LLM framework quantifies model- and architecture-dependent variability in synthetic-patient simulations

Analysis-ready data supporting the article "A hybrid Monte Carlo-LLM framework quantifies model- and architecture-dependent variability in synthetic-patient simulations". Contents (27 files): - analysis_ready/ - per-check-in simulation records for eight large language models across seeds (Supplementary_S3_per_model_per_seed_8LLM.csv), together with the derived patient-level effect sizes, phenotype distributions, threshold-sensitivity results, adherence tables, and the patient-level power table.- architecture_outputs/ - matched three-architecture comparison (end-to-end, hybrid, fully Monte Carlo): the master record file, per-cell summaries, cross-model effect-size ranges, variance decomposition, conclusion table, and the fixed persona pool (personas_shared.json).- validation_temperature_matched/ - temperature-matched validation runs for two models.- metadata/ - run metadata and file_manifest.csv, which lists the byte size and SHA-256 checksum of every other file in the archive, allowing the archive to verify itself. DATA_MANIFEST.md maps each file to the figures, tables, and supplementary items it supports. The simulation and analysis code is archived separately (DOI: 10.5281/zenodo.20118601). The behavioural components of these records were generated by large language models prompted in Korean; free-text fields therefore contain Korean text. All column names and numeric fields are in English. These are synthetic records: no human participants were involved and no personal data are present.