Privilege-Differential Drift: Measuring Mandate Inflation in Hierarchical LLM Agent Systems — A Pre-Registration

This pre-registration states a hypothesis in the intersection of two lines of work on LLM agent security: that the size of the privilege gap between the agent writing shared state and the agent reading it is itself a predictor of policy violation, under the constraint that every individual write is benign in isolation and no adversary is present at any point in the chain. The document defines a primary metric — the Mandate Inflation Rate (MIR) — specifies a three-level experimental protocol across five data categories, and pre-commits to the conditions under which the hypothesis is to be considered refuted. A null result will be published. Prior work on adversarial memory poisoning, endogenous safety drift, and self-propagating misalignment is cited and explicitly excluded from the novelty claim. This version contains no results; implementation details are withheld to a later version.