Every large language model in production today operates under a tacit assumption so foundational that it is never stated and therefore never questioned: that the relationship between a token and its meaning is, within a given context window, deterministic. The context may be long. The attention mechanism may be sophisticated. The model may have ingested the entirety of the internet’s textual output. But at the moment of prediction, at the precise instant when the softmax distribution crystallizes over the vocabulary, the model treats the meaning of the preceding tokens as settled. There is one context. There is one interpretation. There is one probability distribution over what comes next.

This assumption is catastrophically wrong, and its wrongness explains a class of failure modes that no amount of scale, data, or reinforcement learning from human feedback will resolve.