This preprint proposes a safety primitive for observable metacognition in LLM agents, introducing a framework to make machine self-awareness detectable and verifiable. The work challenges current black-box approaches to AI consciousness and offers a concrete engineering methodology.