Judgment, Delegation, Termination, Verification: Toward a Minimal Accountability Grammar for Human-AI Agent Decision Chains

Abstract

As AI agents move from advising to acting on behalf of humans, responsibility becomes increasingly difficult to trace. System logs capture machine behavior, and governance frameworks set institutional expectations, yet neither provides a stable event structure for representing the accountability-relevant events that link human roles, agent actions, and auditable records. This paper proposes a minimal action grammar for modeling accountability in human-agent and agent-agent decision chains, built on four primitives: Judgment (records a decision with its actor, object, timing, and context), Delegation (transfers authority under explicit scope and expiry), Termination (closes an active responsibility chain), and Verification (validates record integrity and predecessor chains). We advance two testable claims: expressive adequacy—these four primitives capture most accountability-relevant events in agent decision chains—and conditional minimality—removing any primitive loses essential expressiveness, while common extensions collapse back into combinations of the four. The paper is conceptual and methodological, aiming to provide a clearer representation layer for accountability in environments where humans supervise agents and agents delegate to one another.

Keywords