The central claim is that learning from scarce observations is not primarily a supervised
classification problem, but an inverse recovery problem over compact functional structures.
In this view, classes are not necessarily predefined labels. They may emerge as stable
compact manifolds or submanifolds in an appropriate representation space. Once these
structures stabilize, expert feedback can assign semantic or physical meaning to them: nor-
mal operation, friction, heat, harmless variation, degradation, emerging fault, or unknown
regime.