Typed Affect Isolation in AI Control Systems: A Formal Separation Between Presentation Adaptation and Epistemic State

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Adaptive AI systems routinely modify tone, verbosity, formatting, rhetorical emphasis, and other presentation properties in response to user interaction. These adjustments are often implemented inside the same stateful runtime that also maintains beliefs, hypotheses, evidence, permissions, revocations, and workflow status. This creates a class of cross-plane failures in which an event intended only to alter presentation can inadvertently modify substantive epistemic or authorization state. This paper introduces a typed affect-isolation architecture in which presentation adaptation is represented as a distinct state plane governed by AFFECT_EVENT and AFFECT_POLICY transition types. Affect events may modify only rendering-related state. They are prohibited from mutating Bayesian belief, hypothesis membership, hypothesis disposition, evidence, provenance, authorization, revocation, lineage, or epistemic thresholds.