Systems that optimize measurable signals can lose access to the goals those signals were chosen to track — not because the signals were poorly designed, but because the optimization process itself removes the feedback pathways the goals depend on. This note identifies signal collapse as a distinct failure class, distinguishes it from governance failure, Goodhart's Law, and overfitting, and documents four instances from the StratoAtlas Series A case archive: differentiation collapse, measurement modality gap, confidence inversion, and population invisibility. A shared mechanism is proposed, and a diagnostic question is formulated. The class remains open. A finite diagnostic procedure is included in the restricted version of this record.
Roman Kir (Mon,) studied this question.