Beyond Normative Monistic Agency: GRACE, a Reasons-Based Neurosymbolic Architecture for Safe and Ethical AI Alignment

The paper proposes GRACE, a reasons-based neurosymbolic constraint architecture that separates an AI agent's normative reasoning from its instrumental decision-making, to constrain agents of almost any design. The architecture has three modules: the moral module determines permissible macro-actions through reasoning based on deontic logic; the decision module encapsulates the target agent and selects instrumentally optimal atomic actions according to the resulting macro-actions; the Guard monitors and enforces moral compliance. The reasons-based formalism provides a semantic foundation for deontic logic, making behavior interpretable, contestable and defensible, and its symbolic representation simultaneously supports formal verification and alignment statistical guarantees enforced by the Guard. The paper uses an LLM psychological counseling assistant as an example to demonstrate how the architecture enables stakeholders to understand, contest and improve agent behavior.