Attractor Geometry and Self-Referential Closure: CDIL Coherence Bounds as Predictors of Behavioral Subspace Separation in Recursive Systems

Read full text Summarise with VIRAT
Authors: Mark Lemon
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Recent empirical work on behavioral probing in large language models reports geometrically separable subspaces in hidden-state representations — regions where self-referential processing and task-directed processing occupy distinct attractor basins detectable before output generation. This paper argues that the CDIL (Cyclic Distinction-Integration Learning) formalism predicts this geometry. Three claims are advanced: (1) the integration tax requires any sufficiently recursive system to partition its state space into self-referential and task-directed subspaces; (2) the coherence bound implies a natural phase transition at approximately 50% of total recursive depth; (3) observed transition depths exceeding 50% indicate coherence failure rather than self-knowledge, reinterpreting the “dark subspace” reported in recent work as a confabulation attractor. Predictions are falsifiable and independent of whether the underlying systems are conscious.

Fields of Study & Subjects

Philosophy and Theoretical Science Complex Systems and Dynamics Psychiatry Mental Health Neuroscience

Publication details

Journal / source
Figshare
Publication date
1 January 2026
Language
EN
Document type
Preprint