🗺️ Embedding Explorer

Each dot is one of the curated Chen et al. (2020) variable stars, placed by a UMAP of the self-supervised encoder's 128-d embedding — the encoder saw only the light curve (flux, uncertainty, timing), never a label. It cleanly separates star classes… but inside each class, tap around: the periods are scrambled — the poster's point. What the map does order smoothly: periodic strictness (how much variance the phase-fold explains) and amplitude — the correlates a naive probe leaks through. The gold ★ is ZTF J1539+5027, the 6.9-minute white-dwarf binary, pushed through the same encoder. Tap a legend chip to isolate a class.