Blind Governance for Autonomous Reasoning

Current Large Language Models fundamentally conflate the distinction between descriptive facts (is) and normative values (ought), treating both as identical probabilistic predictions. This structural homogeneity often results in sycophancy and logical hallucination, rendering monolithic models unreliable for high-stakes autonomous reasoning. In this paper, we argue that robust alignment requires the architectural decoupling of epistemic consistency from normative aggregation. We propose The Rawlsian Architecture framework that segregates reasoning into three isolated stages: a neutral Intake Agent that abstracts user intent; a set of blind Delegate Agents that generate paradox-free options without contextual bias; and a Decision Engine that employs a voting system to select outcomes based on harm minimisation rather than utility maximisation. By enforcing blindness between agents and formalising the aggregation of preferences, this architecture shifts the alignment problem from internal model cognition to verifiable institutional design, offering a neuro-symbolic approach to safe AI governance.