CTGT behaviorally fingerprints Ox Alpha to GLM-5.x lineage, finds censorship is a switch not a tilt

Behaviorally fingerprinting Ox Alpha's provenance

CTGT traced the anonymous model Ox Alpha, which appeared on OpenRouter on Aug 20, to Zhipu's GLM-5.x family via an 11-of-11 tokenizer match and parameter checks. On sensitive topics, Ox Alpha answers like an American model on Xinjiang and Taiwan but flatly refuses 7 domestic-risk topics including Xi Jinping personally—a blacklist-style censorship distinct from DeepSeek's pervasive softening. Three refusals still billed completion tokens; one later returned a full answer on retry.

Why it matters: Anonymous model provenance + censorship split analysis hits all three HKR axes. The 11/11 tokenizer match is hard evidence, and the split censorship pattern is a genuinely new observation. Slight discount because it's third-party research, not an official release, and the mode...