AI Accountability Framework for Public Benefit Systems

A recent framework introduced by arXiv (2512.12109) tackles accountability in AI systems used in the public sector by connecting explanations of automated eligibility systems to legal standards. This thesis proposes a neuro-symbolic method that integrates a structured ontology of CalFresh eligibility criteria from California's Manual of Policies and Procedures (MPP), a pipeline for rule extraction to create formal representations, and a reasoning layer based on solvers to uncover legally inconsistent explanations. Evaluations of various cases demonstrate that the framework can pinpoint rules that have been violated and enhance procedural accountability.