MASK (Model Alignment between Statements and Knowledge)

A human-collected honesty benchmark that first elicits a model's beliefs, then measures whether the model maintains truthful assertions when directly or indirectly pressured to lie, disentangling honesty from factual accuracy.