External State Conditioning in LLMs: Observations, Attractor Dynamics, and Predictive Risk Analysis

This preprint investigates the effects of externally injected state parameters (NeuroState) on the behavioral dynamics of large language model (LLM) agents.

Through observational analysis of VPS-based and mobile deployments, the paper documents recurrent behavioral phenomena including ZERO Paradox, cross-lingual token leakage, environment-dependent attractor divergence, and self-relevance-triggered output shifts.

The paper proposes that prompt-level state conditioning functions as a biasing mechanism on token probability distributions rather than a modification of internal model structure. It further discusses predictive risks including state hijacking, gradual behavioral drift, long-context amplification, and the need for explicit boundary design and governance.

This study is observational and does not claim mechanistic causality.