Verification Allocation in Stylized LLM Agent Pipelines under Hidden Tool Degradation: A Discrete-Time Mean-Field Analysis
Abstract: We study a stylized verification-allocation model for role-structured pipelines with binary inherited corruption and hidden tool degradation. Standard contraction arguments give serial and multi-role fixed points and a finite-population deviation bound under independence. When verification actions affect neither future errors nor observations, the belief-state choice is myopic. We characterize a two-action threshold under an explicit monotonicity condition and a restricted four-action interval and tail policy when generation rates are equal and only one check degrades. Synthetic experiments check these model-specific results and find counterexamples outside the restriction. This is conditional model analysis, not a general mean-field-game result or empirical evidence about deployed LLM systems.