When AI Interns Outperform AI Executives: Role Identity and Information Fidelity in Decision Support

This dataset supports the findings reported in the paper "When AI Interns Outperform AI Executives: Role Identity and Information Fidelity in Decision Support." It contains the scored output of all two thousand four hundred single-hop trials analyzed in the paper, crossing eight organizational personas, three levels of task framing, and four large language models. Also included are the one hundred seed passages with their fact inventories, the eight persona prompt definitions, and the three framing instruction texts used to configure each trial. Each trial routed one seed passage through one role-bearing AI agent under one framing condition. Information fidelity was scored by comparing each AI agent's output against its original seed passage using contextual embedding methods. Together these materials allow reproduction of the results reported in the paper.