The Experiential-Factual Geometric Divide is a Universal Property of Language Representation: Evidence Across 16 Systems

This record contains the full replication package for the paper"The Experiential-Factual Geometric Divide is a Universal Propertyof Language Representation: Evidence Across 16 Systems"(Blaschuk, 2026). We demonstrate that the Experiential-Factual (E-F) geometric divideemerges independently across 16 fundamentally different representationalsystems with LDA cross-validation accuracy 0.866-1.000 (mean 0.957),all permutation p = 0.001 (1000 permutations). Systems tested include7 LLM architectures spanning 2019-2024 (GPT-2-XL, OPT, Falcon,Mistral, Gemma-2, Llama-3.1, Phi-3), a general text corpus (The Pile,sentence-BERT), poetry vs news, clinical narratives vs news,emotional language vs news/ArXiv/Wikipedia, and children's stories(TinyStories) vs news/ArXiv/Wikipedia. Cross-architecture Spearmancorrelation of E-F category profiles yields mean rho = 0.874(range 0.831-0.913), confirming that different LLMs independentlyencode the same geometric structure. Contents: Blaschuk_EF_Universality_2026.pdf — full paper ef_universality_replication.py — end-to-end replication pipeline (Python) ef_universality_results_v3.pkl — full numerical results (16 systems) fig_universality_16systems.png — main figure, all 16 systems fig1_crossarch_heatmap.png — Spearman correlation matrix 7x7 fig2_universality.png — LDA across initial 10 systems fig3_permutation.png — permutation test histograms This work is part of the DSAOP research series and builds directly on: Alieksieienko (2026), DOI: 10.5281/zenodo.19160334 Alieksieienko (2026), DOI: 10.5281/zenodo.19415238 Blaschuk (2026), DOI: 10.5281/zenodo.19455867 Blaschuk (2026), DOI: 10.5281/zenodo.19556416