Human in Control: A Practical Framework for Evaluating Human Control of Advanced AI Agents

This working paper proposes the Human Control Model, a six-dimension framework for evaluating measurable human control of advanced AI agents: Intervention Durability, Technical Containment, Informed Delegation, Oversight Integrity, Control Contagion, and Situational Grounding. It separates behavioral severity from environment reality so that simulations and controlled evaluations are not presented as real-world incidents. Five pilot reviews, HIC-001 through HIC-005, illustrate control-specific measures including detection-to-containment time, permission comprehension gap, monitor independence, and instruction provenance integrity. An additional qualitative stress test informed framework refinement; it is not independent validation. The Meta case is based on a reported user account, with permission history, causal interpretation, and severity remaining provisional. The framework is conceptual and case-based, without inter-rater reliability testing or validated quantitative thresholds. Version 1.0.1 corrects source wording, incorporates the recovered HIC-0 through HIC-5 definitions, and clarifies evidence limitations. Creator: Human in Control Project. License: CC BY 4.0. The deposit includes the PDF working paper and editable DOCX source. Human in Control project website