Original case studies and technical findings, published independently of contracted work — every claim verified against a primary source.
Identified a structural defect in the LLM-as-judge evaluation rubric of a paper published in npj Digital Medicine (9:349, 2026): one exclusionary clause left the evaluator blind to fabricated content while still catching omission. Verified directly against the authors' own supplementary figures. Disclosed privately to the corresponding author with a 14-day window; acknowledged, and publication invited.
A multidisciplinary case study — Psychiatry → Medicine → Neurology — modeling how three services independently reason from an incomplete, partly conflicting record to the same empiric decision to treat, with every lab and imaging figure reconciled against the original source exhibits.
Clinical validation of a 72-rule code-derived rule set powering a synthetic Type 2 diabetes dataset used to train and benchmark medical AI models. A defined five-stage process — direct ICD-10-CM tabular lookup, clinical-correctness review, guideline verification, consequence analysis, severity triage — surfaced 35 defects spanning four independent failure layers: coding validity, clinical categorization, simulation logic, and care-pathway safety.