Artifact-to-Finding Promotion: Evidentiary Requirements for Harm from Correctly Functioning AI Systems

This paper specifies the records required to investigate harm arising from a propositionally accurate AI output whose attributed warrant exceeded its established warrant. It uses artifact-to-finding promotion to name the incident pattern: an AI system performs within specification, no control fails, and a person treats its output as establishing more than it establishes. Existing work separately addresses outcome-graded reliance, contextual interpretation of AI advice, epistemic warrant, and decision provenance. It does not, in the literature reviewed, assemble these elements into an incident-specific reconstruction test for the correct-output case. The paper synthesizes five evidentiary requirements: the output as rendered, qualifiers presented or omitted, the linked decision record, the system's documented warrant at deployment, and contemporaneous model and configuration provenance. Worked successful and failed reconstructions demonstrate their use and the consequence of their absence. The central forensic problem is that the decisive record is the output as rendered to the decision-maker, not merely the underlying transaction log, and conventional logging architectures seldom preserve it.

This paper is issued as a working paper. The incident pattern is offered for testing and criticism rather than as a settled result. Correspondence and challenge are welcome.