Replication package: Large language models as reliable graders of clinical competency for digital standardized patients in a multilingual LMIC

Code and de-identified derived data for the manuscript 'Large language models as reliable graders of clinical competency for digital standardized patients in a multilingual LMIC' (submitted to npj Digital Medicine, 2026). Contents: the grading pipeline (Python; prompt templates for all ten case scenarios, transcript pre-processing, the gateway classification wrapper, run tooling); the analysis (R; pooled leniency GLMM, per-condition Rasch models, agreement and reliability statistics, provider-clustered bootstrap, source language analysis with LLM adjudication, item-level agreement; all manuscript exhibits and in-text numeric macros are generated from this code); de-identified derived grading data (one row per case x coder x checklist item, plus all LLM grading runs by model x language x temperature x repeat); and the full checklist item text for all ten scenarios. Raw DSP encounter transcripts are excluded in all forms due to ethical and privacy restrictions; providers appear only as pseudonymous study codes. See README.md inside the package for the full contents map and reproduction instructions.