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

Sapiens in Control, formerly Human in Control, proposes a six-dimension Human Control Model for reviewing control failures in advanced AI agents. It distinguishes real-world incidents, mixed or contained events, controlled evaluations, and simulations. Version 1.4 builds on v1.3.1 and adds bounded analysis of METR's October 6 transcript-viewer proof of concept and OpenAI's October 6 metagaming research. It proposes separate oversight checks for authoritative records, displayed or exported evidence, reasoning traces, actions, and independently observed outcomes. These are research examples and evaluation proposals, not additional confirmed deployment incidents or completed validation. The six dimensions, five pilot cases, qualitative severity definitions, and prior corrections are preserved. This remains an unreviewed conceptual working paper without completed independent reliability testing. AI tools assisted drafting, source review, and document preparation; this revision is not peer review or private-log verification.