Who Decides the Trade-off? Resolution Policy as Delegation Governance in Autonomous Agents

An empirical study showing that when AI agents face conflicting constraints, today's systems resolve the conflict probabilistically through model sampling—producing outcomes that are unpredictable, irreproducible, and unauditable—and introducing Resolution Policy (the Deterministic Delegation Model) as a formal governance mechanism. DDM makes constraint trade-offs explicit and structurally binding, reducing deviation from 76% to 0% in experiments across two frontier LLMs.

Abstract

When an autonomous AI agent's delegated constraints cannot be simultaneously satisfied, someone must decide which constraint to sacrifice. In current LLM-based agent systems, this decision is made probabilistically by the model's sampling process, producing outcomes that are unpredictable, unreproducible, and unauditable. We term this the Trust Gap. Through 2,248 experimental probes across two frontier LLMs, we demonstrate that a single fallback instruction reduces deviation from 84% to 0%, establishing that behavioral compliance is achievable. However, behavioral compliance is fundamentally distinct from structural guarantee: a single adversarial override reverses compliance from 0% to 100% (R5), and this pattern generalizes across resolution strategies (R7). We formalize the missing element---Resolution Policy---through the Deterministic Delegation Model (DDM): a principal's deterministic, pre-committed trade-off strategy that structurally binds intent to execution outcome. Evaluation across complete 2\times2 factorial designs confirms DDM operates independently of prompt content, injection content, and resolution strategy type. Concurrent work has advanced authorization enforcement; the complementary question---what to do when authorized actions conflict, and by whose authority---is the problem Resolution Policy resolves.