Buffy versus Bella: An archetypometric analysis and comparison

This is a collection of new arXiv papers (July 10, 2026) covering LLM evaluation, AI ethics/governance, data provenance, and domain-specific risks. Key security-relevant findings: (1) "Diagnosing and Repairing Persona Collapse in LLM Advice" shows three leading models collapse >90% of situational advice into a single supportive persona; Inverse-Process Distillation reduced divergence from human persona distributions by ~80% but human raters still often preferred the collapsed default in some contexts. (2) "The Context Access Divide" formalizes a structural inequality where manual context-atten