Foundations / Problem-Naming Series I The Execution-as-Default Problem: A Foundational Failure Mode in Contemporary AI Systems

Description This paper is the first work in the Problem-Naming Series for Relationship-Aware AI. It defines the Execution-as-Default Problem as a foundational failure mode in contemporary AI systems: the implicit architectural assumption that receiving input justifies inference, execution, and output. The central claim is that execution is incorrectly treated as the default consequence of intelligence. As AI systems become persistent, interactive, agentic, and multi-entity, many emerging failures cannot be explained by output quality alone. These include over-execution, premature decision finalization, recursive agent loops, excessive automation, mutual interference, timing misalignment, propagation instability, evaluation inconsistency, and autonomy erosion. These failures arise not because outputs are incorrect, but because execution occurs when it should not occur, occurs too early, propagates too widely, or occurs under insufficiently determined conditions. This paper reframes these issues as failures of execution governance rather than failures of generation, reasoning, planning, alignment, safety, or coordination. To address this, it introduces pre-execution regulation as a necessary architectural principle. AI systems must determine execution eligibility before inference, response generation, tool use, delegation, or action execution proceeds. The paper further introduces the Control Index as a shared precondition for execution-related decision layers, ensuring that execution, timing, intervention, delegation, and propagation decisions do not finalize independently under uncertainty. The key shift is: from output optimization to execution governance,and from action generation to action permission. This work establishes the foundational problem definition for the Relationship-Aware AI research agenda. Its companion foundational paper develops the human-facing manifestation of this problem. Subsequent work extends the agenda toward runtime architecture, evaluation framework, and benchmark protocol.