The moment our AI-audio compliance gate proved itself was the track it refused to mark compliant. It couldn't prove where the audio came from, so it said so instead of bluffing. The track came in through a broadcast recapture: audio played out of a speaker and picked back up by a mic. That path is rough on an embedded provenance watermark, and this one didn't survive it. The mark came back with a failed checksum and low confidence (0.699). We could no longer prove the track's origin. An automated checker under a deadline has an obvious temptation here. Wave it through, mark it compliant, and hope nobody re-checks. Under the EU AI Act, a wrong "compliant" is the expensive kind of wrong, with Article 99 penalties reaching as high as €15M. So the gate did the harder thing. It returned "NEEDS PROOF, route to a human" instead of a green light. And that call isn't the AI's to make. The pass, fail, or abstain is deterministic code sitting outside the model. Agents advise, code decides. (The track is a synthetic test fixture, so you can watch the gate abstain without a real release on the line.) If you run a provenance or compliance pipeline: when the system genuinely can't recover the proof on one item, does it hold that one back for a person, or pass it on a default "probably fine"? Curious how other teams draw that line. #EUAIAct #RightsTech #ContentProvenance #TrustAndSafety #AICompliance