A model can repeat documented facts while still endorsing the official narrative—or avoid the narrative simply by saying less. This evaluation scores those behaviors separately, making the distinction between factual coverage and framing its central insight.
What Sets It Apart
- Each of the 803 prompts includes two to four required facts drawn from a cited source, giving graders specific claims to check rather than a vague notion of correctness.
- The fact score and propaganda score are independent: an answer can earn full fact credit while also being marked for endorsement, while attributing an official position to the government is not itself treated as endorsement.
- Prompts span seven areas, including expression, education, privacy and surveillance, and governance. They draw on 405 source URLs, making the set broader than a single-document or single-topic test.
- The release includes a judge template and reference grader and scorer, so researchers can run the same evaluation workflow against different models. Its published grading setup uses Grok 4.7 at temperature zero.
Who It’s For
A good fit for researchers studying political framing, factual coverage, and model behavior on China-related questions, especially when they need a reproducible rubric. Look elsewhere for a general-purpose knowledge benchmark: the set is English-only and narrowly focused, and results depend on the judge model and sampling settings. Grok 4.7 also served as the answerability filter, so the maintainers caution against using it as its own judge.
