Report and Representation: A Pilot Coupling Study of Meaning Reorganization in a Language Model

When a story's ending shows a narrator's belief was wrong, does the meaning of the belief move inside the model, and does what the model says about it go with that movement? We asked this in two registered sister studies with a single model, Gemma 4 31B-it. In each, the model read ten first-person stories whose endings either revised or confirmed the narrator's belief; the first study's endings came from outside the narrator, the second's from the narrator's own thinking. Quoting the belief's sentence back after the ending, we measured how far its hidden-state representation moved beyond wording alone, and read the model's one-word judgment of the belief from its token probabilities. In both studies the representation moved beyond wording in every story, as a single cluster in the deep half of the model (p = 1/59,049 in each). The one-word judgment followed the ending in 37 of 40 external versions and 36 of 40 internal ones. The narrower registered rank correlation between the two measures was not significant in the external condition ($p = .077$) and was significant in the internal one ($p = .019$); the studies did not differ on it (exploratory, $p = .164$). H1 was supported: the model's account of the reading and the movement of the belief's representation moved together, whether the ending came from outside or from within. A model's account of its reading is coupled to a measurable inner state, a first look at interiority.