Capability Primitives and Their Risk Surfaces
Explore how capability primitives transform generative AI system functions and alter trust boundaries, risk profiles, and evidence requirements. Learn to assess discrete capabilities from text generation to model adaptation, understand their dominant failure modes, and identify necessary controls and evaluations for regulatory compliance and operational governance.
The postmortem template from the last chapter raises a key review question: what capabilities and processing steps did the system use behind the chat interface?
A capability primitive is the smallest architectural capability that changes what a GenAI system can do. Adding a capability primitive can also change the system’s trust boundaries and evidence requirements. In a regulated environment, the user experience can look similar even when the underlying system uses different capability primitives. Review the underlying capabilities: is the system only generating text, or is it retrieving governed data, executing actions, or using sensitive interaction data for logging, evaluation, or model improvement? ...