Authority Leases and Concurrent Evidence Failure in Safe Agentic Execution

Abstract

This bounded technical report introduces authority leases for safe agentic execution: execution contracts that must still hold when a model, worker, tool, scheduler, or delegated automation acts. The report uses a frozen masked high-priority review set, fail-closed measurement checks, synthetic negative controls, and local proxy replay to describe a local class of action-time evidence-validity failures. It does not claim population-level probability, general failure rates, reviewer reliability, production validation, or customer outcome improvement.

Agentic systems often make a decision before they act. A model, worker, automation, or delegated tool can inspect evidence, queue an action, and execute later. Between decision and action, identity, permission, policy, target state, availability, actor, impact, or timing can change.

This paper argues that a recurring class of AI workflow failures is not best explained as generic reasoning error alone. In this class, the system acts on evidence whose freshness, identity, authority, visibility, or materialization timing no longer matches the action being taken.