The standard mental model of an LLM embedding space treats it as Rn, but symbolic descriptions of points—token strings, labels, prompts, or programmatic names—come from countable languages. This creates a basic modeling question: if the ambient embedding space is made measure-theoretically explicit as (Rn, B (Rn), λn), what Lebesgue-measurable part of it can be pinned down pointwise by a countable symbolic system? We show by a short packing argument that the answer is null: for any partial map L: Rn ⇀ Y with Y countable, the set of points locally isolated within their own fiber is at most countable, hence Lebesgue-null. The result is a diagnostic of the Lebesgue idealization, not a claim about empirical embedding studies, which use finite vocabularies, activation distributions, and other non-Lebesgue measures. We complement the theorem with an empirical typology of cl100kbase showing eight token types that are invisible to the ambient Lebesgue model, and close with three alternative formulations of the underlying question.
Zhongmang Cheng (Wed,) studied this question.