AI Delegation Study: Honest Reporting Leaves Surplus Unclaimed Across 5 LLMs

A new theoretical paper from Taksch Dube (arXiv 2607.14357) tackles a fundamental question in AI alignment: when should you describe your true intentions to your automated proxy? The work, submitted July 2026, introduces "within-range regret" – a single quantity that determines whether honest reporting is optimal. Theorem 1 shows honesty is optimal exactly when the proxy is loyal, meaning it already plays the best action it can reach. This formalizes the intuition behind prompt-engineering and jailbreaking: safety constraints that shift a model's behavior while keeping its best output reachable make honest descriptions suboptimal, so sharper reports can gain.

The paper's central result is the "Trilemma of Aligned Delegation" (Theorem 2): no guardrail can be at once binding (displacing the truthful action from the proxy's best reachable outcome), truthful (honest reporting stays optimal), and capability-preserving (that best outcome remains reachable through some report). Any two preclude the third. The authors estimate within-range regret from samples and track it across model updates at a cost proportional to drift, not update frequency. When tested on production LLMs from five major providers under an alignment-style cap, they found honest reporting consistently leaves surplus unclaimed – recoverable by inflating the user's report. The 31-page paper includes 7 figures and 4 tables, bridging game theory, multiagent systems, and theoretical economics.