Confirmatory phase: materials, data and code for "Specifying large language model-based virtual patients for clinical psychology simulation"

Confirmatory phase of a computational study on how the specification of large language model-based virtual patients for clinical psychology simulation (case format, symptom level and model) affects their consistency between sessions, control and specificity. Four written cases, each in three formats and at three symptom levels, were played by three open models run locally (Gemma 3 27B, Qwen 3.5 27B, Mistral Small 3.2 24B; Ollama 0.34.3). The confirmatory phase tested nine hypotheses from the exploratory phase (https://doi.org/10.5281/zenodo.23037676) with 3162 new sessions (38430 requests) in eight arms: replication of questionnaire sessions and interviews, questionnaire sessions with the number of symptom days removed, cross-questionnaire sessions, reworded interviews, and no-case controls. The analysis plan, hypotheses and analysis programs were fixed by SHA-256 hash before generation; the plan was not registered publicly. The record contains the generation and analysis code, the configuration, the plan and incident log, all session records (except Study 1 records of OCI-R administrations, which contain copyrighted items; their numeric item answers are included), item-level responses, analysis outputs and the figure program. See README.md for the file list and instructions.