The middle column is the evidence layer — every image's strongest signal (priority: a face (recurring or tagged) → camera EXIF → document → appearance-only). Hover a signal for what it means. Ribbon width ∝ image count; the split into personal / reference is the verdict.
Read it: a face (recurring person or a tagged one) is the strongest signal — 5,759, almost all personal (tagged → a hard 0.95 floor); camera leans personal; appearance-only images (no face/camera/tag — 31,078) are the ambiguous bulk the probe splits mostly to reference. These interpretable evidence signals are the cold-start; the shipped verdict is the supervised probe on the SigLIP embedding, which agrees with them ~86% and decides the rest.