ElevenLabs watermark: 25 of 81 reads still in progress - the public detector caps daily checks. Every row is backed by the public corpus: originals, attacked copies, hashes, commands, verbatim detector responses. A disputed result gets rerun against the corpus.

Provenance and interoperability checks above; all durability results in the battery below - label gone on every attack, watermark not testable (no public interface; the strongest open mark dies to a 5% speed change and streaming compression). Evidence A, B and C archived in the corpus with raw detector responses. 1 of 4 checks passed.

Provenance checks above; watermark-on-every-path and all durability results in the battery below - intact through container, codec, dynamics and edits, gone on tempo, pitch, or the acoustic path. The public detector errored for one window (10:27-10:28, recovered by 11:11); its classifier channel is speech-oriented, the watermark channel is the music verdict. 1 of 2 checks passed.

Columns are the provenance layers a third party can actually check: C2PA manifest via the c2pa reference SDK, watermark via the vendor public detector. Suno exposes no public watermark interface, so that column is not testable. FAIL on a C2PA column means the manifest is gone; every attacked copy still plays. Suno C2PA column measured on corpus track s01 (Sep 24, 2026); ElevenLabs watermark column measured on the 002 corpus; NT cells are queued, corpus expansion in progress. Raw results: corpus.

Suno exposes no public watermark interface, so only its label column is measured. ElevenLabs watermark reads still landing - gray turns green or red as the daily detector cap allows.

Decision line 0.5. Scores from the authors' released SpecTTTra-alpha-120s model, middle-chunk inference per their code. The detector was trained on 2024-25 Suno and Udio output; every AI track above is current-generation and un-attacked. Control: the human recording is also clean in the ElevenLabs watermark check (no mark found, 15:56 PDT). One open detector, not every detector. Per-file scores archived in the corpus.

Attacking does not hide a track from this detector. Most attacks leave it dead; a few wake it up - trimming an ElevenLabs track is what finally flags it. The numbers move both ways, which is itself the finding: detection confidence is not a stable signal.

FULL = every 10-second chunk of the payload recovered, PARTIAL = some chunks, DEAD = none. The ElevenLabs column shows only attacks already read through its public detector (about one read per hour); the rest are queued. One track, 40-bit payload, github.com/sony/silentcipher. Raw matrix: corpus.