Sapiens in Control: A Six-Dimension Framework for Evaluating Control Failures in Advanced AI Agents

Abstract

Sapiens in Control, formerly Human in Control, is the same independent public-interest research initiative and working-paper series. This paper proposes the Human Control Model: Intervention Durability, Technical Containment, Informed Delegation, Oversight Integrity, Control Contagion, and Situational Grounding. Version 1.3 changes the project and paper name while retaining the six dimensions, five pilot cases, qualitative severity definitions, prior versions, and concept DOI lineage. It adds analysis of OpenAI's September 28 safety-case guidance, Anthropic's September 24 Project Swap experiment, and Agent Incident Registry research. New proposals distinguish preference accuracy from authorization and connect control claims to bounded evidence. These sources are not additional uncontrolled deployment incidents. This is an unreviewed conceptual working paper, not an independently validated framework. Earlier versions retain the Human in Control title. Revision notes explain the transition. Version 1.3.1 is an audit correction: it explicitly includes the disputed Muse permission account, identifies model-judged evaluation frequencies and simulated trusted labor, and qualifies proposed measures and empirical independence. The original v1.3 remains available at https://zenodo.org/records/23132853. This audit is not peer review or private-log verification.