The Value of What the Model Knows — paper outline

This research outline sets out a programme for studying governed knowledge as a variable in the economics of AI tasks. It proposes to connect the knowledge supplied to a language agent, including claim content, review status, permitted use and access stratum, with task success, human review, refusal, escalation and the cost of maintaining that knowledge. The proposed paper examines the relationship between classical value-of-information theory and the behaviour of language agents and human reviewers. It investigates the conditions under which governed knowledge improves net workflow value, recognising that additional context can reduce performance and that the set of economically feasible knowledge states may be empty. The economic analysis includes the upkeep of reusable claims and the costs associated with refusing or referring an act. The proposed experimental design separates changes to claim content through actual review, the visibility of review labels, and enforcement through a status check. It specifies independent outcome adjudication, measurement of review and escalation costs, controls for leakage and drift, and a programme of public and enterprise workflow evaluation. Publication status: This deposit is a research outline, not a completed or peer-reviewed paper or a report of experimental results. Its hypotheses, proposed derivations and measurement design remain subject to formal development and empirical testing. Cordier’s simulation layer has not been independently rerun; the outline states the assumptions on which its proposed analysis depends.