Behaviorally Fingerprinting Ox Alpha's Provenance and Censorship
Summary
Analysis of Ox Alpha, a GLM-family model seen on OpenRouter, focusing on behavioral fingerprinting, censorship patterns, and provenance. The report uses LineageEval to compare censorship across topics, revealing a bimodal suppression on China-related prompts and a reliance on a blacklist for sensitive content. The findings align with GLM-5.x characteristics and include token-level investigations and prompts designed to uncover model provenance.