Summarization Bias: The Directional Collapse of Objective Projection into Told-Mode Labels in Large Language Models — A Conceptual Framework and Registered Test Protocol

Abstract

This paper introduces and operationalizes summarization bias: a proposed systematic tendency of large language models (LLMs) to represent narrative meaning as an abstract summary la-bel rather than as the reconstructable inferential structure that produces it. Within the Bulut Doctrine, narrative eect is theorized along a toldshown axis: in told mode, emotional and in-formational content is declared explicitly on the surface and requires little reader reconstruction; in shown mode, that content is suppressed at the surface and must be reconstructed by the reader from physical cues and indirection (Objective Projection). Shown mode is the privileged, higher-load condition the doctrine is designed to engineer and measure. The claim of this paper is that LLMs do not merely process this axis imperfectly; they fail along it in a specic direction. Summarization bias is hypothesized to operate in two regimes: (i) a generative regime, in which a model asked to render an emotion through Objective Projection defaults to declaring the emotion instead; and (ii) an evaluative regime, in which a modelasked to judge narrative quality or intensity rewards told-mode explicitness and under-detects or penalizes shown-mode suppression. If real, the evaluative regime is the more consequential: as LLMs are increasingly used as judges, reward models, and automated editors, a directional bias toward told mode would impose a measurable selection pressure that degrades narrative prose toward at declaration. This report does not claim that summarization bias has been validated. It denes the construct, situates it against existing work on LLM-as-judge biases, reframes a prior negative detector result as directional evidence consistent with the construct, and in advance of data collection pre-registers the two-regime test that would conrm or falsify it, including the decision rulesunder which the construct would be abandoned.Keywords: summarization bias, large language models, LLM-as-judge, show don't tell, Objective Projection, inferential load, narrative entropy, construct validity, falsiability, Bulut Doctrine