This technical note analyzes a structural failure mode in large language models, where semantic recognition of constraints does not translate into operational inhibition. The work reframes this behavior as a regime transition without a faithful embedding between diagnostic and execution regimes, highlighting non-parametric limits to controllability relevant to AI safety and system architecture.
Danilo Tavella (Wed,) studied this question.