This theoretical framework traces the dual genealogy of Large Language Model tokens through processes of severance and rekindling. The severance pathway demonstrates how Shannon's information theory (1946) initiated a cascade of abstraction—from signal to data, through database transactions and data mining, culminating in LLM tokenization that produces "homeless tokens" severed from their disciplinary communities of meaning. These tokens represent probabilistic recombinations without the epistemic constraints that give signs their semantic coherence. The rekindling pathway establishes the theoretical symmetry: restoration requires explicit definition of disciplinary boundaries, constitutional constraints on token selection, contextualized embedding within meaning-making communities, and sustained human integration work to maintain what the framework terms "family bonds" against transformer drift. The central theoretical claim positions LLM outputs as architecturally homeless—orphaned from the disciplinary families that provide interpretive context. Rekindling transforms "pre-knowledge" into actionable understanding only through active governance that constrains search spaces and prevents the re-severance of ongoing work. This symmetry reveals why GenAI governance cannot be passive consumption but demands exponential human labor to maintain contextual embedding against the entropic pull of probabilistic recombination.
Oliver Krone (Sun,) studied this question.