APR-SafetyNet v0.1 RFC Specification: Epistemic Governance for AI Systems Answering US Public Benefits Questions

Abstract

APR-SafetyNet establishes a substrate-layer governance framework for AI systems interacting with US public benefits programs. It addresses predictable epistemic failure modes — especially program confusion (e.g., Medicare vs Medicaid) — and defines mandatory behaviors for clarification, disambiguation, and safe procedural guidance. This RFC defines a canonical data schema for program knowledge, disambiguation rules for high-confusion program pairs, a governance constitution that all implementations must obey, and a minimal JSON/YAML bundle representing APR-SafetyNet v0.1. Intended audience: Public-interest technologists, benefits navigators, AI governance architects, safety-net practitioners. This specification is motivated by the documented epistemic failure mode in which AI systems provide confident procedural guidance based on misidentified benefit programs — causing real harm to the most vulnerable populations by delivering authoritative answers to the wrong question.

Ähnliche Arbeiten