Somewhere, Not Nowhere: First-Person Register, Epistemic Limitation, and Perspective Formation in Language Models

Language models learn about experience almost entirely from third-person description: text that reports what happened, what things are, and what people feel, written from outside the experience it describes. This paper asks whether a different training signal — synthetic first-person experiential narratives — can form better functional models of perspective: the situated, indexical, action-oriented point of view from which an agent perceives partially, acts under constraint, errs, and updates. We distinguish description from perspective and argue that third-person corpora, however large, may under-determine perspective-taking, because the point of view is precisely what objective description leaves out. We propose a corpus of first-person narratives structured by perceptual limitation, goal, obstacle, action, error, and revision, varied across agent types, including non-human umwelten, as a distinct pretraining or fine-tuning signal, together with a content-matched control that crosses register (first- vs third-person) with epistemic limitation (a marked, partial vantage vs an omniscient one), so that the grammatical voice and the represented vantage can be told apart rather than confounded. The claim is functional, not phenomenal: the corpus is not proposed to give a model experience or to make it conscious, but to test whether first-person register, represented epistemic limitation, or their interaction teach a measurable skill (perspective-taking, deictic reasoning, self/other distinction, agency attribution, and point-of-view-dependent decision) that an omniscient third-person rendering does not. We locate the proposal among synthetic-corpus, persona, egocentric, and embodied-experience work; differentiate it from first-person datasets built for interpretability or classification; and position it against the interaction-only prescription of the "era of experience." The predicted signature is a dissociation: register- or limitation-driven training improves perspective-dependent tasks without improving general factual recall. We give the dataset design, the evaluation, and the conditions under which the proposal is wrong.