Tracks whose full-track embedding is near-identical (cosine ≥ 0.995) — the same
recording regardless of how the metadata reads. These are the duplicates the id / slug / exact-metadata signals
miss: cross-artist collab reuploads, compilation copies, differing credit strings, re-masters. Listen to each pair and
mark it — verdicts save in your browser; Export when done.
⚠️ ≥0.995 is not automatically a duplicate. It reliably finds the same recording under messy
metadata (different artist credit, title typos, comp reuploads) — but a track's radio edit and extended mix
also score this high while being genuinely different releases. The tell is duration: a big length gap (flagged
Δ…s) with a Radio / Extended mix label = keep both; near-equal
length = almost always the same file. When unsure, listen.
–candidate groups
–random tracks scanned
–had a ≥0.995 twin
0you've judged
Pure-vector scan over a random sample · verdicts stored locally in localStorage · no merges happen here.