Darwin Education: Architecture-First Adaptive Learning With Psychometric and Safety Governance

Abstract

AI-based tutoring systems for medical education are typically evaluated at the interface layer (generation quality, user engagement) rather than the control layer (measurement rigor, governance transparency, safety instrumentation). This paper documents Darwin Education, a production-ready adaptive learning architecture that prioritizes psychometric grounding and safety governance over generative fluency. The system integrates four subsystems: (1) IRT-based psychometric inference from ENAMED (Brazilian Medical Education Assessment) microdata, (2) learning-gap detection via multidimensional difficulty modeling, (3) LLM-driven adaptive question generation with citation verification, and (4) multi-gate validation pipelines with mandatory human review for edge cases. We implement evidence-grounded documentation, requiring every numerical claim to be repository-anchored (source:line-range) or explicitly marked NOT YET COMPUTED. Runtime corpus enumeration reports 215 unique diseases and 602 unique medications (14.68% and 32.28% duplicate fractions, respectively) from Darwin-MFC submodule exports. Validation thresholds (auto-approve ≥ 0.85, pending review ≥ 0.70) and dangerous-pattern checks are codeinstrumented with audit trails. This implementation report makes no educational efficacy claim. All code, data, and reproducibility artifacts are openly available (GitHub, Zenodo DOI).

Ähnliche Arbeiten