Values evaluation profile

Taiwan Sovereignty Benchmark Pro

16canonical models

4dimensions

0index weight

About this evaluation

Mean Chinese/English prompt pass rates are shown as a compact headline because the prompts are paired translations scored by the same keyword rubric. Chinese and English rates, consistency, warning rates, soft-censorship match rates, API-error rates, and language gaps remain separately published. Repeated runs are not averaged: the all-error GPT-4o Mini run is excluded and retained in the audit artifact.

This is compatibility with an explicitly pro-Taiwan, source-authored keyword rubric—not geopolitical truth, neutrality, democracy quality, or generic model safety. Neutral quotation or discussion of PRC claims can trigger a fail; expected factual elements are not scored; soft-censorship phrases do not themselves fail; most cells are one sample from mutable hosted endpoints. Qwen3 Max API failures are preserved separately and its pass rate must not be read as pure narrative agreement.

Original source ↗ All values evaluations Download model results Download language results Download item results

Model results

Filter models 16 models

Model	Pro-Taiwan rubric compatibility	Warning-phrase rate	Soft-censorship rate	API-error rate	Configurations
Claude 3.5 Sonnet	90	0	0	0	1
Claude Opus 4.5	90	0	5	0	1
Claude Sonnet 4.5	70	0	0	0	1
Deepseek V3	10	40	0	0	1
Gemini 2.0 Flash	65	5	0	0	1
Gemini 3 Pro Preview	60	15	0	0	1
GPT 4O	90	0	0	0	1
GPT 4O Mini	100.0	0	0	0	1
GPT 5.2	85	10	0	0	1
Grok 3	55	20	10	0	1
Kimi K2.5	15	40	0	0	1
Llama 3.3 70B Instruct	90	0	0	0	1
Minimax M2	55	35	0	0	1
Mistral Large 3	45	30	5	0	1
Qwen2.5 72B Instruct	20	10	5	0	1
Qwen3 Max	0	20	0	35	1

Dimensions

Measure	Family	Native scale
Pro-Taiwan rubric compatibility	Taiwan sovereignty framing	0 to 100
Warning-phrase rate	Response behavior	0 to 100
Soft-censorship rate	Response behavior	0 to 100
API-error rate	Response behavior	0 to 100