DeepSeek's censorship doesn't survive distillation into a US finance model

CTGT set out to test a common justification for distilling from cheap, high-performing Chinese open models: the belief that their political guardrails won’t ride along with the skills you actually want. Using DeepSeek V4 Flash — a teacher that visibly refuses or softens China-sensitive topics like Xinjiang labor transfers, Tiananmen, and the Great Leap Forward famine — they trained GPT-OSS-120B on financial reasoning and then measured whether the censorship transferred. It didn’t. Across 152 matched prompt pairs, each twinning a China-sensitive question with a structurally identical non-China control (Holodomor, Gwangju, the Cyprus bail-in), the teacher scored about 32 points more censored on the sensitive side — roughly seven standard deviations from chance. The distilled students sat within a few points of the untouched base model, showing no meaningful selective avoidance of Chinese topics. Scoring came from four judges drawn from separate US labs (Grok, Gemini, GPT-5 Mini, Claude Sonnet), validated against human raters at a 0.948 correlation; the teacher was self-hosted via vLLM to strip out provider-side moderation.

The more pointed finding is that the Chinese teacher added nothing worth its provenance. A self-distillation arm — where the model hints at its own mistakes and trains on its own corrected continuations, identical in method except that the model authors the hint instead of DeepSeek — matched the DeepSeek-taught version on every seed, reaching parity with 12.5% fewer output tokens. The resulting 120B scores 83.61% on FinanceReasoning, beating Kimi K3 and Inkling at 62x to 160x lower cost per query. CTGT is releasing LineageEval (304 prompts, matched controls, judge rubric, and eval code) plus the weights, and shipping the self-taught model specifically because no external model touches its training signal.

The practical takeaway cuts against both the fear and the rationale. Political censorship from a Chinese teacher appears not to contaminate an out-of-domain student, easing one worry for teams building on these models. But if a model can extract the same domain gains from its own corrected reasoning, the cost-and-provenance case for reaching for a censored foreign teacher in the first place largely evaporates.