The AI System as Closed-System Test Bed is a methods paper by Nobel Glas that translates the Semantic Deviation Principle into quantities measurable from language-model probabilities.
The first quantity compares each realized token’s surprisal with the model’s expected surprisal at that position. Positive values indicate a token less expected than the model’s baseline; negative values indicate convergence toward highly probable continuation. A second, more expensive quantity compares sampled future continuations with and without a contextual intervention.
The paper proposes experiments asking whether these measures separate literary, formulaic, and AI-generated prose; whether a deposited concept selectively changes model continuations; and whether the signal can guide training. The Crimson Hexagonal Archive corpus is treated separately to reduce self-validation.
The work is a pre-registration rather than a finding. “Meaning,” “slop,” and “deviation” remain operational terms inside the experiment, not quantities already shown to correspond to literary or human significance.