Large language models frequently generate outputs that are locally coherent while remaining globally semantically unstable, resulting in hallucinations and meaning drift. Prevailing mitigation strategies rely on threshold-based metrics, guardrails, or post hoc alignment mechanisms, which scale poorly and impose externally defined normative constraints. This paper introduces a frequency-based stabilization principle that treats meaning as a dynamic relational process sustained over generative time, rather than as a static correctness condition. Semantic stability is not enforced through predefined thresholds or truth criteria, but emerges through the persistence and recurrence of coherent relational configurations during generation.
biljana knezevic (Thu,) studied this question.