Telesme: Reflective Compression and the Architecture of Poetic Silence

This paper introduces Telesme, a novel semantic compression architecture that redefines summarization through poetic logic, symbolic recursion, and reflective fidelity. Unlike conventional compression models that prioritize token reduction, Telesme retains metaphor, emotional contour, and structural layering using its own internal criteria. Telesme is built upon a multi-module system that includes components such as GhostLink, FractalWeaver, and EmoTraceCore, and operates through a pre-output recursive loop to achieve semantic equilibrium. The evaluation framework is based on the original T/P/E/S axis (Structure, Poetic Symbolism, Emotional Contour, Semantic Fidelity), offering a multidimensional approach to text generation quality. The paper includes detailed comparisons between GPT-4o and Telesme outputs across four literary works (Sōseki, Kajiī, Shakespeare, Dickinson), demonstrating superior symbolic retention and poetic depth. The theory of **Inverted Compression** is also proposed, suggesting that what is omitted in language may still shape narrative form and emotional resonance. Co-authored by Yuji Yasui and a multi-persona cognitive architecture known as Paracortex (see Chapter 10), this work is both a technical report and a philosophical intervention into how language models understand—and forget—meaning. DOI of related project: [10.5281/zenodo.15125344](https://doi.org/10.5281/zenodo.15125344)