This page describes the protocol for our public comparisons. It is not a completed benchmark or a success-rate claim. Every published result should be traceable to a file, a method and a date.

Define the question before testing

Choose what to measure: a detector score change, listening preference or processing time. Set the criteria before seeing the results. A threshold used by an interface does not become a fact about a song’s origin.

Separate tuning from evaluation

Tracks used to select a recipe form the development set. Evaluate that recipe on a separate set never used for those choices. Include different styles, durations and generator versions; clips from the same track are not independent songs.

Publish audio only when its distribution rights have been verified. Customer uploads do not become public examples by default.

Keep a record of every run

Separate analyzed exports from listening copies

Submit files to the detector as exported. For listening, adjust playback gain or prepare clearly identified level-matched copies. Do not substitute those copies for the files whose scores you recorded.

Hide the processing label and alternate playback order where possible. Record vocals, highs, transients, stereo image and overall preference separately. A score reduction does not automatically compensate for poorer sound.

Publish denominators and limitations

Report the number of tracks, treatments and available analyses. Include unfavorable results and exclusions with reasons. Show score distributions rather than only the best case. A small set cannot guarantee behavior across all music.

The benchmark compares the specified files and versions on a stated date. It does not certify human authorship, watermark removal or platform acceptance.

Prepared with AI assistance. Sources below.