Cedar (Clinical Evidence‑Driven Agent for Risk‑prediction) is a four‑stage pipeline that analyzes structured electronic health records, social‑determinants data from clinical notes, and, for type‑1 diabetes, continuous glucose monitoring traces. It publishes a probability of a future outcome, a table of cited evidence, and a written reasoning chain. In a retrospective study of 2,489 type‑2 and 4,018 type‑1 paediatric patients, Cedar achieved a pooled area‑under‑the‑receiver‑operator‑characteristic curve (ROC‑AUC) of 0.815 in type‑2 and 0.756 in type‑1 diabetes, surpassing classical machine‑learning models (0.700 and 0.631, respectively) and temporal models (0.656 and 0.603). For eleven cohort‑outcome pairs, Cedar ranked higher in nine cases, with confidence intervals that separated for dyslipidaemia in both cohorts. Only four of 70,246 evidence citations did not match the chart, and two pediatric endocrinologists reached Cedar’s predictions in nine of ten cases. The study was single‑center, retrospective, and the sample sizes were limited to the two cohorts, so generalisation to other settings or broader populations has not yet been established.