Uniformity Asymmetry: An Exploratory Metric for Detecting Representational Preferences in LLM Embeddings – Code and Dataset

Replication package for the paper "Uniformity Asymmetry: An Exploratory Metric for Detecting Representational Preferences in LLM Embeddings". CITE ASD'Elia, D. (2025). Uniformity Asymmetry – Code and Dataset. Zenodo. https://doi.org/10.5281/zenodo.18110161 CONTENTS• uniformity_asymmetry_clean.py – Standalone Python validation script• Uniformity_Asymmetry_Validation.ipynb – Google Colab notebook• dataset.json – 230 statement pairs across 6 categories• Validation results for 4 models (Gemma-2-9B, Llama-3.1-8B, Mistral-7B-v0.1, Apertus-8B)• Bitcoin blockchain timestamp (.ots) for priority proof METHODUniformity Asymmetry measures representational clustering differences between semantically equivalent statements with different framings. Side A contains abstract/conceptual formulations, Side B contains concrete/numeric formulations. Statistical validation uses 10,000 bootstrap resamples with 95% confidence intervals and Cohen's d effect sizes. IMPORTANT CAVEATThe dataset design introduces a structural confound: Side A statements are consistently more abstract/conceptual than Side B. Observed asymmetries may reflect representational compression rather than normative preferences. These findings are exploratory and require validation with balanced datasets. DISCLAIMERThis is an independent research project conducted by the author in a private capacity. The affiliated institution provided no funding, supervision, or resources for this work. REQUIREMENTSPython 3.10+, transformers, torch, accelerate, numpy, scipy LICENSEMIT License