buk81/uniformity-asymmetry: v4.0 — Paper 4: Alignment Robustness

Paper 4: Alignment Robustness Depends More on Training than Architecture DOI: 10.5281/zenodo.18316488 A systematic cross-vendor study examining how preference optimization (RLHF/DPO) affects attention head specialization across 8 vendor families and 25+ model variants. Key Findings | Finding | Evidence | |---------|----------| | SI Reduction Pattern | RLHF/DPO reduces SI in unprotected architectures | | SWA Protection | Sliding Window Attention correlates with SI preservation | | GQA Noise Sensitivity | ~5,800× higher PPL-slope than MHA | | Synthetic Immunity | Phi SI ≈ 0.33 across 10.8× scale | | Training Override | Qwen2 shows no recursive degradation | Robustness Hierarchy: Training Methodology > SWA > Architecture > Scale Reproducibility Package 19 A-Tier Notebooks organized by claim (A1-A5) 27 JSON Result Files with standardized schema claims.yaml — Machine-readable claim mapping verify_release.py — Package integrity checker reproduce.py — CLI reproduction tool Quick Start cd paper4 python verify_release.py 📄 Full Paper (PDF)