Capability Space Structural Decomposition

Structural Problem

AI capabilities are typically discussed as a single dimension — "more capable" or "less capable." The structural problem is that capability is not one-dimensional. It decomposes into orthogonal axes — autonomy (the degree of independent action), generality (the breadth of applicable domains), and intelligence (the depth of reasoning and problem-solving) — that can be developed independently and that couple to safety in different ways.

A system can be highly autonomous but narrow, highly general but not autonomous, or deeply intelligent but constrained. Each combination creates different risk profiles and safety requirements. Treating capability as a single dimension collapses these distinctions and prevents effective capability steering and safety boundary management.

System Context

This application operates in the AI capability assessment and development planning space, addressing how capabilities are measured, how they relate to safety, and how development decisions affect the capability-safety landscape. The relevant system boundary includes the multi-dimensional capability space, the safety regions within that space, and the development trajectories that navigate through it.

Diagnostic Capability

• Capability space decomposition mapping an AI system's capabilities onto orthogonal axes to reveal its structural position
• Safety region mapping identifying which regions of the capability space are structurally safe, conditionally safe, or high-risk
• Development trajectory analysis predicting how planned capability improvements will move the system through the capability-safety space
• Coupling analysis identifying how gains on one capability axis affect safety properties on other axes

Typical Failure Modes

• Capability-safety decoupling where capability development proceeds without tracking the safety implications across all axes
• Axis coupling surprise where improving one capability dimension unexpectedly degrades safety on another dimension
• Safety region exit where incremental capability improvements cross a safety boundary without detection because the boundary exists in a different dimension than the improvement

Example Use Cases

• Capability development planning: Structural analysis of how proposed capability improvements affect the system's position in the capability-safety space
• Safety assessment framework: Providing a multi-dimensional capability map for comprehensive safety evaluation
• Regulatory compliance mapping: Translating capability-safety analysis into regulatory frameworks that distinguish between different types of capability

Strategic Relevance

AI governance and regulation increasingly require nuanced understanding of AI capabilities. A structural decomposition of capability space provides the analytical foundation for governance frameworks that are calibrated to actual risk rather than simplistic capability thresholds, enabling both more effective regulation and more productive capability development.

SORT Structural Lens

The SORT framework addresses this application through four structural dimensions, each providing a distinct analytical layer.