Tag
This paper demonstrates that formal robustness certificates for embedded neural interface models can pass even when task accuracy collapses under adversarial attack, and proposes a unified empirical audit framework to address alignment failures between training objectives and operational user welfare.