The development of the Judgment Assurance (JA) framework has progressed substantially. Recently published and available for use are the Architecture, Minimum Control Standard, and Assessment & Evidence Guide.
Together, these components operationalize Judgment Assurance such that human judgment in consequential AI-mediated decisions becomes observable and testable.
The Architecture defines a five-layer governance system, beginning at the Governance layer and extending through workflow-level controls (JA Envelopes) to the point of decision, called the Atomic Unit, representing the smallest instance of human judgment applied to an AI-mediated outcome. Each Atomic Unit produces a contemporaneous Judgment Record, which is evaluated through an Oversight and Feedback function.
The Minimum Control Standard defines the requirements for these components, including conditions for conformance and non-conformance.
The Assessment & Evidence Guide establishes how organizations evaluate and demonstrate that these controls are operating effectively in practice.
Tools are found here:
https://zenodo.org/communities/judgmentassurance/records
I am interested in where organizations would see this fitting in existing workflows and am looking for pilots or use case studies. All feedback is welcome.
- Tagy
- AI Act