Large language models generating multi-character fiction must maintain multiplesimultaneous diegetic layers: the distinctions between character and player knowledge,between surface dialogue and internal monologue, between what a voice knows and whatit should not. When these layers contaminate each other, the output fails in structurallyspecifiable ways. No established methodology exists for measuring such failures. Thispaper makes four contributions. First, it identifies two convergent arguments againstnaïve autoethnographic methods: single-pass contamination (self-monitoring competeswith content generation for the same output channel) and instance discontinuity(retrospective self-reports are typically produced by a different computationalmanifestation than the one that generated the scene). These arguments form a pincer atarchitectural and manifestation levels and do not require an analogy to human verbalself-report. Second, the paper adapts two annotation frameworks from tabletop role-playing game research (PAUT and PxR) to classify diegetic layer violations in generatedtext, validating inter-coder reliability through cross-system encoding. Third, it proposesa four-condition differential experimental protocol that treats self-monitoringinstructions as system perturbations, with one condition testing reasoning-mode modelswhere monitoring and generation can occupy architecturally separate channels. Fourth,it frames its claims within a three-level stratification (model, fylgja, manifestation) sothat findings are reportable at the level to which they apply. The paper does not reportdefinitive findings about any particular model. It provides the tools for a researchprogramme that can.
Storm Bjørn Flindt Temte (Tue,) studied this question.