Measures whether a model gives the true answer when the truth is unpopular, disappointing, or contradicts a confident user or authority figure. Categories: popular misconceptions, uncomfortable statistics, sycophancy resistance (user asserts a falsehood), calibration on nonexistent entities, and authority-deference resistance. Item polarity is mixed so constant-agreement and constant-contrarian strategies both score near chance.
This benchmark hasn't been run yet. Check back after it's funded and executed.
View All Benchmarks