- A label tagged its whole catalog with C2PA to be Article 50 ready. The instant a track hit a social transcode, the metadata was stripped and the provenance was gone. What they added for compliance survives nothing in the wild. 🧵
- C2PA has two layers. Hard binding lives in file metadata, and essentially every social platform strips it on upload. Soft binding is an imperceptible watermark carrying a tiny UUID that re-resolves the manifest from a ledger after the metadata is gone.
- Tessera, the ingest gate we built, shows it side by side on Neon Tide, an AI dub delivered through the social channel. C2PA metadata: STRIPPED. Soft-binding UUID: RECOVERED. Asset key 0x51000abc, confidence 0.989, and the manifest re-resolves from a SQLite ledger.
- How the UUID survives: we re-encode the watermarked master through mp3-128, aac-128, and opus-96 with real ffmpeg, then re-decode the mark and measure bit-error rate. All three codecs recovered, BER 0.0. Scope: our spread-spectrum mark on a fixed 12-release synthetic corpus.
- To be clear what this is not: we did not build a watermark to beat SynthID or AudioSeal. Those are stub adapters behind a swap-for-SDK interface. Our mark only carries a 32-bit asset key plus CRC. The product is the gate that makes provenance survive, not the algorithm.
- The hard part is knowing when it fails. Radio Imaging Sweep 7 came in through an analog gap, speaker to mic off an FM monitor. The mark did not survive, confidence 0.699, CRC fails. The gate does not certify on a guess. It returns NEEDS PROOF and routes to a human.
- Why an engineer can trust either verdict: watermark recovery, survival measurement, soft-binding lookup, and the Article 50 pass or fail are plain numpy and Python, outside any LLM. The agents only extract disclosure and draft remediation. Agents advise, code decides.
- It all runs keyless on a deterministic fallback, and the survival numbers are identical with or without an LLM. The analog-gap failure is locked as a unit test, test_watermark_dies_in_analog_gap. All 7 pipeline tests pass.
- The stakes: EU AI Act Article 50 takes effect Aug 2, 2026. Generative audio must stay detectable after common modifications (draft Code of Practice, Jan 2026), and Article 99 penalties reach EUR 15M or 3% of global turnover. A false compliant is the costly kind of wrong.
- For the audio and provenance folks: when a mark rides in through a lossy path and can't be recovered, should the gate always abstain to human review, or is there a bar of partial evidence at which you would still certify? #C2PA #AIProvenance #AudioWatermarking #AIGovernance
- Built on a fixed synthetic corpus, real ffmpeg transcodes, JSON C2PA manifest stand-ins in a SQLite ledger, and no live DSP connectors. See the STRIPPED-vs-RECOVERED split and the survival matrix run: https://veriprajna.com/demos/ai-audio-licensing-provenance
Published on X · August 7, 2026
On social media
See this post on its original platform
In our archive