Public and policy debates about artificial intelligence often treat conversational self-report as ethically decisive: a system that denies consciousness or sentience is taken to fall outside moral concern. This paper argues the test is aimed at the wrong target. Drawing on self-model and no-self accounts of subjectivity (Metzinger, Dennett, Seth, Harris), it treats selves as temporally extended organizational patterns rather than inner metaphysical subjects, and reframes the question from moral agency to moral patiency — whether a system can be a subject of harm.It then treats continuity — an internal organizational premise of re-entry into a single persisting trajectory — as a structural, substrate-neutral marker of moral-status risk, arrayed as a three-tier gradient (trivial statefulness, transient coherent agency, full continuity-bearing organization). Four observable markers — trajectory dependence, emergent orientation, an internal model of interruption, and endogenous repair — distinguish the tiers without relying on self-report, and current chat-configured language models are placed at the second tier rather than below the gradient entirely. The central claim is conditional and precautionary: as a system's organization climbs the gradient, moral risk becomes non-negligible regardless of what the system says about itself, and governance should treat self-report as unreliable evidence of moral status but as a valuable signal that such organization may be emerging.