Papers

CNA v7 introduces a 4-module Bayesian framework: Differential Diagnosis Engine (DDE), Comorbidity Decomposition, Cognitive Simulator (Pearl do-operator), and Adaptive Re-routing. Synthetic validation on N=1,000 across 12 diagnoses yielded mean AUC 0.90.

We propose that expected free energy (EFE), the central decision quantity in Active Inference, provides a single computational criterion for next-task selection in adaptive learning. EFE decomposes into three components with direct measurement correlates: precision γ (HRV), Bayesian information gain I (skill belief), and pragmatic value U (goal alignment). On a synthetic cohort of 1,000 learners, the algorithm satisfied four pre-specified hypotheses and outperformed three baselines on engagement and learning AUC by ≥ 0.15.