AudioTrust Reconciles C2PA Provenance and AudioSeal Watermarks

AudioTrust is a local verifier that reads two independent trust signals on AI-generated audio - C2PA provenance metadata and the inaudible AudioSeal watermark - and reconciles them into an auditable verdict. It doesn't create or sign anything; it only inspects existing evidence and reports trusted, partial, unverifiable, or contradiction (an "Integrity Clash" between a human-origin claim and a detected AI watermark).

Two design choices stand out. The tool never relies on AudioSeal's decoded message, since it proved non-reproducible across test runs; only the stable detect_prob score is used. Origin classification comes from the manifest's action claims (e.g. generatedBy, softwareAgent) rather than the source_type field, since real C2PA signatures from c2patool typically omit source_type entirely.

An independent audit on a clean clone caught a serious logic bug: recordings with realistic human device names (e.g. "Zoom H4n") were being misclassified as synthetic by default, making the contradiction verdict - the tool's core value proposition - nearly unreachable. The fix changes the default to indeterminate unless explicit synthetic or human-capture evidence exists.

The project followed a spike-then-spec-then-audit workflow, with issues logged transparently rather than hidden behind passing tests. One gap remains open: a fully real end-to-end fixture for the trusted/contradiction path is still pending, blocked by a c2patool WAV-signing limitation, and tracked as KI-5.