Hallucination as Enactment: The Invisible Assumption in Large Language Models
Year: 2026
Type: preprint
Abstract: Hallucination in large language models is defined on the output. It is found afterwards, by whoever checks. This paper says the cause comes earlier. Before a model answers, it decides for itself what the input means and what the task is. That decision is an assumption. Nobody sees it when it is made, and every next step builds on it. In Context Psychology this first step is called an enactment. A language model makes it itself, without a fixed outcome. A hallucination is the assumption that later turns out to be wrong. A correct answer rests on the same kind of assumption, which happened to be right. The paper reviews the published research. The parts are known: models infer a hidden task, give all text the same priority, guess rather than say they do not know, commit before they answer, and do not show what they used. The step that connects these parts was not found. Six testable hypotheses follow from the thesis, and the studies found refute none of them. All current measures work on or after the output, or inside the model. None of them fixes the outcome outside the model before the model acts. The paper ends with a practical way of working with a language model, based on the EOCME cycle. Theory: The Holographic Meaning Field, doi.org/10.5281/zenodo.19385072 EOCME® by Noemeos. Patent pending: EP 25 212 132.2 and PCT/IB2026/061076
Author: Blokker, Robert, Constant
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236637656
