BiblioEngine: A Stylometric Transformation Framework Utilizing LLM Hallucinations for Generative Art

Title: BiblioEngine: A Stylometric Transformation Framework Utilizing LLM Hallucinations for Generative Art

Abstract: BiblioEngine is a text analysis engine designed to transform literary texts into reusable numerical parameter sets by quantifying scenic density, narrative tension, the distribution of silence, and the interpretative fluctuations inherent in Large Language Models (LLMs). The primary objective of this project is to redefine literary works not merely as objects of passive appreciation, but as high-dimensional "Digital Assets" operable within digital art and generative expression.

The system performs line-by-line analysis of literary texts. Notably, BiblioEngine intentionally executes multiple analysis iterations on identical text segments to extract the non-deterministic variations of LLM outputs. These variations are captured and stored as a specific parameter set termed "ghosts." This design architecture enables the inevitable ambiguity of literary interpretation to be treated not as computational noise, but as a manipulatable creative variable.

Methodology Brief: The inference engine employs Google's Gemini 1.5 Flash, configured with high temperature (T=0.95) and Top-K sampling (K=40) to amplify interpretative fluctuations. A distinguishing feature is the "Ghost Parameter" (P_ghost). The system executes N trials for the input. For each parameter v, the standard deviation sigma(v) is calculated. The aggregate is defined as:

P_ghost = (1/|V|) * sum(sigma(v))

This value represents the "richness of ambiguity."

Output: This paper presents the complete algorithm specification of BiblioEngine v1 and the structure of the resulting 14-column CSV data.

References: [1] BiblioEngine Algorithm Specification v1.0 (2025) [2] Plutchik, R. (1980). "A general psychoevolutionary theory of emotion" [3] Moretti, F. (2013). "Distant Reading" [4] Bender, E. M., et al. (2021). "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" [5] Coelho, P., et al. (2017). "Quantitative Analysis of 27 Emotions in Berkeley Study"