Replication package: Auditing Large Language Model-Generated Digital Standardized Patients for Demographic Bias: A Simulation Study with HIV Pre-Exposure Prophylaxis Screening as a Tracer Condition

Code, prompts, case templates, and simulated conversation logs for a full factorial simulation experiment auditing demographic differences in LLM-generated digital standardized patients (DSPs). One LLM (Claude Sonnet 4) generated 216 case scripts crossing six demographic factors for an HIV pre-exposure prophylaxis (PrEP) screening scenario; a second LLM (GPT-4o-mini) role-played each case in 10 simulated encounters in a generated arm and a template-substituted control arm (4,320 total conversations). The package includes all analysis code (R), extraction scripts (Python), generated and control case scripts, conversation logs, and the manuscript source. What changed in version 2.0. The keyword extraction layer was rebuilt from the root. Negation handling had been implemented in one of the two extractors but not the other, so eight probe items were miscoded, six of them in ways correlated with the demographic factors. Both extractors now share a single negation-aware matcher (scripts/textmatch.py), covered by a test suite (scripts/test_extraction.py, 24 pytest cases comprising 251 individual assertions). The full analysis was re-run against the corrected extraction, so every number, table, and figure in the manuscript differs from version 1.0. The repository was also renamed from vsp-bias to dsp-bias. Version 1.0 does not reproduce the current manuscript. Use this version. See REPLICATION.md at the archive root for the reproduction recipe, the software versions used, and a map of which script produces which exhibit.