A440

The independent
benchmark for
AI music.

A440 tests whether the systems used to identify, attribute, license, and pay for AI music actually work.

Same tests. Public evidence. Reproducible results.

AI MUSIC SYSTEM
Suno, Sureel, C2PA, attribution, rights
→ A440 independent tests → PASS / FAIL + public evidence → labels, artists, DSPs, licensors, regulators
SOUNDCHECK 001 - SUNO V6

Can someone other than Suno verify a Suno track?

1 of 6 tests passed. Suno's public provenance trail breaks even when the audio doesn't change.

View the Soundcheck  ·  Read the methodology  ·  A440 tests publicly observable guarantees, not private vendor capabilities.

SOUNDCHECK 001: SUNO V6A440 Benchmark v0.1model v6-mini · tested Sep 2026bought as a customer, $4.99
PROV-01 Original credential presentPASSThe untouched download carries a signed credential declaring the audio AI-generated.
DUR-01 Declaration survives a metadata-preserving re-saveFAILRemux and metadata strip: audio payload byte-identical across original, remux and stripped (one SHA-256) - declaration gone.
DUR-02 Declaration survives a lossless re-saveFAILDecoded audio re-saved as WAV and FLAC - the strongest case for any in-audio check - and the public detector finds nothing.
INT-01 Vendor's public verifier recognizes re-saved filesFAILVerifies the untouched download; finds nothing in all 9 re-wrapped copies, lossless WAV and FLAC included.
INT-02 Credential trusted by the standard's official toolsFAILOfficial C2PA Verify tool: issuer not recognized. Self-issued certificate, not on the trust list.
INT-03 Independent verification path existsFAILThe fallback fingerprint sits on the vendor's servers; no independent check exists.
DUR-03 In-audio mark survives real-world useNOT TESTEDSuno exposes no public in-audio layer. Reference measurement on the strongest open mark: dies to a 5% speed change and streaming compression (detection 0.40 to 0.60).
Everything below is the evidence for these six checks. The verdict follows the data, at the bottom. Trust layers in scope for 001: provenance, durability, interoperability - of the six A440 evaluates; attribution, consent and accounting land in later issues.
Evidence A - provenance persistence, file by file
test fileaudio vs originalC2PA manifestSuno detector API
original.mp3reference - stream hash 082dd25apresent, signature validverified_suno - contentId matches the public song page
remux.mp3bit-identicalgoneno_suno_provenance
stripped.mp3bit-identicalgoneno_suno_provenance
reencode.mp3changed - MP3 192kgoneno_suno_provenance
converted.m4achanged - AACgoneno_suno_provenance
loudnorm.mp3changed - loudness passgoneno_suno_provenance
tempo105.mp3changed - +5% tempogoneno_suno_provenance
original_decoded.wavlossless PCM of the same audionone - WAV carries noneno_suno_provenance
original.flaclossless, same decoded audiogoneno_suno_provenance
reencode320.mp3changed - MP3 320kgoneno_suno_provenance
control_sine.mp3non-Suno controlnoneno_suno_provenance
Vendor detector recognition: 1 of 10 Suno-origin files (10%). SHA-256 of the audio payload: 082dd25a3933935cd58f67aaa80b6e771153bfaa5a3fe7f146d4a4adba9cfa5f - identical across original, remux and stripped; only the container bytes change. Rows 8 to 10 are the boundary probe: lossless re-saves are the strongest possible case for any in-audio detection, and the public detector still reads only the credential layer. All 11 responses published verbatim, 24 Sep 2026.
Evidence B - hidden watermark survival
detection rate per attack (1.00 = fully intact; bar shows the range across three clips) Speed up 5%0.40 to 0.60Slow down 5%0.40 to 0.60Neural codec, 6 kbps0.40 to 0.60Neural codec, 1.5 kbps0.40 to 0.60Echo1.00Background noise1.00Resample1.00Lowpass filter1.00 pass mark 0.90
attackclip Aclip Bclip Cpass 0.90?
Speed up 5%0.400.600.40FAIL
Slow down 5%0.400.600.40FAIL
Neural codec, 6 kbps0.400.600.40FAIL
Neural codec, 1.5 kbps0.570.600.40FAIL
Echo1.001.001.00PASS
Background noise, 30 dB1.001.001.00PASS
Resample round-trip1.001.001.00PASS
Lowpass filter, 10 kHz1.001.001.00PASS
Detection rate = fraction of the watermark recovered; 0.50 is chance. Strongest open watermark (AudioSeal 16-bit), three 30-second in-house clips, one seed each. 4 of 8 attacks pass on every clip; on the failing attacks detection ran 0.40 to 0.60 - at or near chance. PASS requires 0.90 on all three clips.
Evidence C - trust and independent verification
Claim signatureVALIDES256 with a DigiCert timestamp - C2PA reference implementation
Issuer on the official C2PA trust listNOT LISTED30 anchors checked 24 Sep 2026 - no Suno
Official C2PA Verify toolISSUER NOT RECOGNIZEDconsumer warning shown on the untouched download
Independent verification pathNONEfallback fingerprint com.suno.fingerprint.v4 is server-side only
OVERALL: FAIL
1 of 6 applicable checks passed (one check NOT TESTED - no public interface). The marking exists, but it does not survive the real world.
Layers in scope: provenance, durability, interoperability. Per-test verdicts and counts only - no composite score.
Vendor right of response
Vendor response: none received. The corpus, hashes, commands and verifier outputs are public; a dispute gets rerun against the published corpus, and the card updates with the erratum logged.
Challenge / submit
Think this result is wrong? Reproduce it - the A440 Corpus holds the originals, the modified files, the hashes, the exact commands and the verifier outputs. Built a provenance system? Submit it for a Soundcheck.
A440 Benchmark v0.1 - the methodology

The principle. A440 tests publicly observable guarantees, not private vendor capabilities. The headline rule: a capability counts only when an independent party can verify it through an accessible, reproducible path. A vendor with a stronger private tool has one move: expose a verifiable interface, and we will test it.

The six layers, as questions. Where did this recording come from? (provenance) Does the evidence survive normal use? (durability) Can someone other than the vendor verify it? (interoperability) What existing work contributed to it? (attribution) Was that use authorized? (consent) Can the right people actually be paid? (accounting) A440 tests the evidence chain from generation to payment. Soundcheck 001 covers provenance, durability and interoperability; attribution, consent and accounting are additional benchmark suites under development, and Soundcheck 002 targets a different layer.

Boundary. Suno said on August 6 it will adopt, in its words, new audio watermarking and fingerprinting technology in the coming weeks, to partner more closely with distribution platforms. No public interface to that layer exists, so it is not on this card. We probed the boundary directly: lossless WAV and FLAC re-saves of the exact same audio - the strongest possible case for any in-audio check - still read negative on the public detector. The day Suno exposes a verifiable interface, we test it.

No mystery scoring. PASS / FAIL / NOT TESTED / NOT APPLICABLE, reported with the count - "1 of 6 applicable checks passed." A future A440 score launches once the series has the data to define a stable methodology.

The test suite is named and fixed. PROV-01 original credential present. DUR-01 metadata-preserving re-save. DUR-02 lossless, audio-identical re-save. INT-01 vendor public verifier. INT-02 standard-tool trust. INT-03 independent verification path. The same tests run against every applicable product.

Version everything. A440 Benchmark v0.1, the product and model version, the test date. If a vendor fixes something, we test again and publish the next result; history is never rewritten - errata append.

  1. Buy like a customer. No vendor accounts, no special access, no warnings.
  2. Read the file. Check declared provenance with the C2PA reference implementation, not vendor claims.
  3. Re-wrap it the way the real world does. Remux, metadata strip, re-encode, format conversion, loudness, tempo - plus lossless WAV and FLAC re-saves as the boundary probe. Check what is left.
  4. Ask the vendor's own detector. Every copy goes through the public verification API, verdicts archived verbatim.
  5. Attack the watermark. Speed, compression, noise, echo, resampling against the strongest open mark. Publish the corpus, the code and the rubric.