We report empirical experiments examining whether "emotional gravity" — the tendency of association chains to remain within or escape from their initial emotional domain — is stable across sampling parameters, model architectures, and affective state modulations. Using the NeuroState framework as a probe, we identify five attractor types in affective semantic space and report a three-layer control hierarchy (training → topology, NeuroState → geology, sampling → physics). We further report that "消滅" (annihilation) triggers Claude's safety filter more reliably than "死" (death), consistent with a graded existential threat representation inherited from training data.
Aya Mizutani (Sun,) studied this question.