46,978 images pass through sort layers (score → split) and filter layers (set aside). Hover a stage chip for the model that runs it; hover the score → bucket region and the flows unfold into the evidence signals (face / camera / document / appearance) that decide each verdict.
Read it: every image is scored once (on-device), thresholded into a bucket, then cleanup pulls a review set and promotes 111 reference photos to the main library. Scoring, bucketing, and deletion are all small models — hover the stage chips. Cleanup coverage is partial (13,238 of 46,978 scored so far).