Artifact Observatory
Measured field reports about AI-generated music: what detectors actually catch, which delivery formats change the evidence, and which restoration choices remove artifacts without erasing the music.
Comparisons use the same files and preprocessing wherever the claim requires it.
Every report states its partition, threshold, date and important caveats.
Public datasets, result tables and methodology are linked at the claim.
Eight Public AI-Music Detectors on One Test Set
A controlled 2,104-file comparison, why F1 alone hides false-positive behavior, and an audible A/B of restoration as a separate downstream task.
Publication contract
Reports are snapshots, not timeless product claims. A new model, dataset partition or correction gets a new reviewed date and a change note. Submitted audio is never published or indexed without explicit permission; the first reports use Intrect-controlled or already-public research material.