AI safety has advanced through a wide range of research, frameworks, standards, regulations, technologies, and implementations across both institutional and technical domains. On the institutional side, these include laws and regulations, international frameworks, standards, risk management, organizational governance, audit, accountability, and incident reporting. On the technical side, they include alignment, interpretability, evaluation, robustness, agent safety, runtime monitoring, runtime enforcement, and provenance.

These efforts have developed to address different problems, and each constitutes an important element of AI safety. This Architecture does not begin from the premise that these existing efforts are inadequate. Rather, it assumes that many strong “threads” already exist.

At the same time, there is not necessarily a consistent structure for connecting institutional requirements to specific AI decisions and executions within organizations, or for determining the institutional basis and Authority under which technical safety and control mechanisms are applied.

The issue, therefore, is not the absence of strong threads, but how those threads are connected.