Auditing proprietary alignment in large language models: a comparative framework without ground truth

Two researchers propose a statistical framework for auditing proprietary alignment in large language models under black-box access, that is, specific policies that vendors inject without official disclosure. The paper's background is existing reports that some models' responses reflect proprietary rules and organizational interests, causing censorship or misinformation on controversial topics. The framework quantifies systematic deviations between the target model's responses and those of a set of baseline models in a shared semantic space, using relative behavioral divergence rather than absolute correctness as the basis for auditing. The authors state that they have applied it to several previously widely discussed but unquantified cases, providing a systematic and scalable foundation for external evaluation of vendor-specific alignment behavior.