buk81/uniformity-asymmetry: v1.1-preliminary: Extended Results - Output Correlation Analysis

Extended Results: Output Correlation Analysis Status: Post-paper validation experiments (2026-01-01) This release adds validation experiments testing whether Uniformity Asymmetry (UA) correlates with model output preferences. --- Key Discoveries Gemma r=0.95 is an artifact GT-Numeric category drives 95% of the correlation Embedding-Output Disconnect Llama: neutral embeddings, biased outputs Multilingual inversion in Apertus Concept compression vs. output precision Lie vs Bullshit classification Categories cluster by confidence × asymmetry --- Highlights Gemma: r=0.953 → r=-0.608 when removing GT-Numeric Apertus: r=-0.07 → r=+0.587 when removing GT-Numeric Llama: Neutral UA but strong output preference --- Added Files FINDINGS_SUMMARY.md — Executive Summary Output_Correlation_Experiment.ipynb — Reproducible Colab data/ — Raw analysis data figures/ — Visualizations provenance/ — Bitcoin timestamp proof --- Relationship to Paper These results complement the main paper without contradicting it. The paper's cautious "exploratory" framing was validated. ---Provenance SHA256: 5520612f4ba370c098caf954f9a03a1f324cdb39ffe102461acfd57a48ad7bfb Timestamp: 2026-01-01 17:00 UTC