Consistency over accuracy: run-to-run stability of contemporary large language models on Turkish curriculum-aligned theoretical anatomy multiple-choice questions

Stability across repeated administrations is essential for educational use of large language models (LLMs), yet it is rarely quantified in non-English, curriculum-aligned anatomy contexts.

Eleven contemporary LLMs answered 100 Turkish, faculty-authored, curriculum-aligned anatomy multiple-choice questions from AYDEP, targeting the undergraduate Physiotherapy and Rehabilitation anatomy curriculum in three independent runs (≥ 12-hour intervals). Testing used developers' web interfaces in Turkey with browsing disabled and default generation settings (August-September 2025). Performance was summarized with a stability-aware 0-3 item score (number of correct responses across three runs) and predefined response-consistency classes.

A subset of models achieved near-ceiling totals with high run-to-run stability, whereas others showed greater session-to-session variability (i.e., changes in the selected option across independent runs initiated as separate sessions under identical inputs). Several nominal differences among higher-performing systems did not remain significant after multiplicity control. Within-family updates produced selective, not universal, gains. Many models exhibited medians of 3 (IQR 3-3) on the 0-3 scale (ceiling effects), and lower means were accompanied by larger dispersion. Consistency profiles provided information beyond mean accuracy by distinguishing reliably correct from volatile behavior. In addition, we observed "consistent & wrong" patterns on a subset of items, where the same incorrect option was repeatedly selected across runs.

In Turkish, curriculum-aligned anatomy items, contemporary LLMs can be both accurate and stable, but single-trial accuracy can mask volatility and stable systematic errors. Adoption decisions should prioritize stability-aware appraisal (including consistent-correct and consistent-wrong rates), with local validation on institutional item banks and periodic re-evaluation as models evolve. Extending this framework to multimodal anatomy and constructed-response tasks will further inform trustworthy, learner-facing use.