Tagging accuracy

Dry-run scans every photo/video in a date range and checks how well the tagger would have done, without changing any tags. It uses whichever reference and negative photos are already set up on the main enrollment page (this page has no refs of its own); edit refs there and re-run here to see the effect. Pick a date range and detection confidence below, run once, then tune thresholds live with the explorer further down, no re-scan needed for that part. Only mismatches are shown as thumbnails; correct matches are counted in the table but not shown.

This evaluation compares against your existing Immich tags in that range, so it's only as accurate as those tags are. Before running it, make sure every photo/video in the range has already been manually tagged and double-checked, otherwise the results below will be misleading.

Thumbnail badge: a number like "YOLO 0.68" means the detector found the pet at that confidence; "fallback" means the detector failed and the whole photo was scored instead.