Tag
This paper develops the QQ equality from cognitive science into an audit criterion for LLMs, characterizes theoretical mechanism classes, and empirically tests on an open-weight instruction-tuned model, finding that saturation (near-deterministic responses) prevents distribution-level audits.