Hansa Dhammahaso

The paper proposes an advanced auditing framework for classical-to-modern LLM translations, demonstrating that embedding drift signals potential error severity rather than error itself, and identifying model-specific failure modes in high-drift translations.