Compliance systems built on large language models typically assume that the verdict they return is shaped by the regulatory rule they are given. A new arXiv preprint (arXiv:2610.12313) tests that assumption directly, asking whether five models actually change their compliance decisions when the rule changes across 20 regulatory and platform-policy domains.
Rather than judging whether a verdict is correct, the study treats rule sensitivity as the object of audit: if a verdict does not track the rule, the system is not really complying with the specified policy. The available abstract lists examples such as "delete" and "swap" before cutting off, suggesting the test covers concrete policy actions.
Because this is a single preprint, there are no independent sources to compare. The excerpt does not include the study's numerical results, so this digest describes the research question and method rather than reporting final findings.