Four verdicts, on screen
Every release is synthetic: short procedurally generated tones, no real music, artists, labels, or takedown owners. The names are fictitious test fixtures. The numbers are the live engine's output, re-run 2026-08-07.
The batch, and the operating number
Twelve releases queue for distribution, and the gate returns 7 COMPLIANT, 2 routed to remediation, and 3 BLOCKED, a 58 percent auto-clear. That is the number a Head of Trust and Safety funds against: we cleared 7 of 12 automatically and flagged the rest with named owners. The header carries the market context that makes it urgent, Article 50 effective Aug 2, 2026 and penalties up to 15M euro or 3 percent of global turnover (EU AI Act), and the full-batch regulator dossier exports as HTML or JSON.
Hard binding stripped, soft binding recovered
Neon Tide (REL-001) is an in-house-AI release delivered through the social channel. The panel shows the whole thesis side by side: hard binding via C2PA metadata reads STRIPPED on transcode, while soft binding via the watermark UUID reads RECOVERED at confidence 0.989, with the recovered asset key on screen and the manifest re-resolved from the ledger. The survival matrix re-encodes the master through mp3-128, aac-128, and opus-96 with real ffmpeg, and the UUID re-decodes at bit-error rate 0.0 on all three. DDEX disclosure is complete and a takedown owner is present, so the gate returns COMPLIANT on evidence under Article 50(2). This is the single most important real-world fact about C2PA, shown rather than asserted.
A recovered mark is not a clean bill
Glasshouse (REL-003) recovers its watermark at confidence 0.992, with BER 0.0 across all three codecs, so provenance is fine. The gate still returns BLOCKED, under Article 50(4): the DDEX AI-disclosure is incomplete, missing instrumentation, mixing, and mastering because the aggregator fixture did not pass them through. Shipping as-is would be an active Article 50 disclosure violation, so the track routes to a named owner with a 5 day SLA. Mark present is not the same as compliant, and the gate checks the whole chain rather than stopping at the watermark.
When the mark does not survive, it abstains
Radio Imaging Sweep 7 (REL-004) arrived through the analog gap, recaptured speaker-to-mic off an FM monitor. The mark did not survive: soft binding is Not recoverable at confidence 0.699 and the CRC fails, so the gate returns NEEDS_REMEDIATION, NEEDS PROOF, routed to human review, do not certify. The honest nuance is right there on screen: the survival matrix run on the clean master shows the mark would survive normal transcodes, but this track came through the analog-gap channel that actually broke it, and the gate refuses to bluff. That abstention is encoded as a unit-tested invariant, not a hope.
The regulator dossier, per track
Every one of the 12 tracks exports to a dossier, HTML and JSON, carrying the verdict and cited clause, the recovered UUID and manifest id, the survival matrix with BER per codec, DDEX completeness, the remediation owner and SLA, and the adversarial reviewer's sign-off or downgrade. Solstice Engine (REL-009) shows a subtle case: the mark recovers at confidence 0.991, but no C2PA manifest was ever registered in the ledger, a process gap, so it routes to remediation with the fix spelled out. Coverage is complete, with every track carrying a cited-clause verdict and nothing left unexplained.
