The full path for all 46,978 images: the evidence signal that scores each (hover a signal), the personal/reference bucket, and the cleanup outcome. Ribbon width ∝ image count.
Read it: evidence signals (a face merges recurring-person + tagged) score every image; the bucket split is the verdict; cleanup pulls 7,872 to review, promotes 111 reference photos to the main library, and leaves the rest. The signals are the interpretable cold-start; the supervised probe on the SigLIP embedding is the final arbiter (~86% agreement).