Contemporary large language models exhibit a characteristic pattern of failure under sustained semantic ambiguity: initially fluent responses degrade into inconsistency, blending, or post hoc correction as competing interpretations accumulate. These failures are typically attributed to hallucination, reasoning error, or insufficient context. In this note, I argue that such explanations misidentify the source of instability. Through simulated testing, I show that the failure arises from the absence of a basic representational capability: an explicit mechanism for prioritizing meaning when interpretations compete. Introducing even a minimal external constraint that forces unequal weighting among interpretations produces immediate and systematic improvements in observed response stability. This suggests that current architectures lack a primitive required for semantic stability and that many observed failure modes are downstream effects of this absence rather than defects in reasoning or knowledge.