Textual Identifiability and Anchored Semantics for Evaluating World-Model Claims in Large Language Models
Abstract
Claims that large language models (LLMs) learn world models often conflate decodability, causal employment, and external anchoring. This Theory article separates these claims and makes their evidential requirements computable. A text-only learner is represented as a Markov kernel: equal corpus laws imply equal learned-state laws, while corpus and training-procedure discrepancies yield explicit discrimination bounds on a declared finite horizon. For internal representations, a conditional localization model identifies a target-loading subspace, and a nullspace construction determines whether a target can be changed while specified associated readouts remain invariant. Rank, stability, and intervention-coverage conditions delimit both positive and negative causal conclusions. Anchored structural semantics records inferential behavior, structurally admissible interpretations, and independently evidenced world links. A complete two-switch example enumerates the interpretation class, computes its structural quotient, and translates linguistic or visual records into constraints on a fixed model. Its accepted records reduce eight target denotations to two and then one, while leaving irrelevant parametrization unresolved. A separate reproducible Pythia-14M application measures decodability and controlled causal sensitivity on eight held-out prompts; a crossed filesystem-count audit supplies no accepted external anchor. These bounded demonstrations do not constitute empirical validation of the general framework, which remains future work. The contribution is a conditional operational framework that preserves the distinction between internal causal evidence and external reference while allowing partial, cumulative anchoring evidence to be reported.