Answer-Materiality as the Regulated Quantity for Refreshing Exogenous Knowledge: A Preregistered Measurement Protocol and Interface Sketch

Systems that hold knowledge whose changes originate outside the system must decide how often to re-check. The scheduling side of this problem has an established theory: refresh under a finite budget is formalised as a restless multi-armed bandit and solved via Whittle indices, with optimality statements (Whittle 1988; Age of Information after Kaul et al.; Age of Incorrect Information; fresh caching after Abolhassani, Tadrous, Eryilmaz and co-authors; Koley and Singh 2024). These policies regulate on a change signal, such as the age of version. For knowledge content, that signal carries less than its name suggests: Mansoor, Ahmad and Yoon (2026) report that of 396 content changes detected by hashing on open-web pages, 34.3 percent affected the correctness of a cached answer, ranging from 4.5 to 67.0 percent across five freshness classes; there the figure serves as a correction factor in the evaluation rather than as the regulated quantity. At the same time, deriving materiality from content shows limited separability in reported measurements (AUROC 0.59 for separating contradiction from repetition; 55.2 percent for recognising invalidated memories). This record preregisters a measurement protocol for the answer-materiality rate of knowledge-block classes in agent architectures, a secondary policy simulation that swaps only the regulated quantity under an equal checking budget, and an interface sketch in which a knowledge block carries validity, checking duty and materiality as three separate fields. No results are reported; the protocol is fixed prior to measurement. SHA-256 hashes of the four measurement question sets are included in the document, evidencing that the instruments were fixed before any change data were inspected.