Your deepfake detector cleared this claim photo. It was right. The photo is real. That is exactly why it was fraud. In building our Claims Evidence Firewall we staged a synthetic auto claim around the oldest trick in the book: a genuine camera photo lifted from a settled prior claim and re-submitted as new damage. A single-signal authenticity check waves it through, because there is nothing fake to detect. So we stopped asking one question ("is this image fake?") and ran a forensic crew instead. Provenance checked out, and the pixel-forensics signal (a simulated stand-in for a learned artifact model) had nothing to flag, since the photo genuinely is real. The catch came from two signals a lone detector never runs: the Reuse Analyst matched its perceptual hash to a settled prior claim, and the Narrative-Consistency Analyst saw the account contradict the photo. 🔎 The referral itself is not the model's call. A deterministic policy gate reads the typed findings and routes the claim to SIU. Agents advise, code decides, so the outcome is reproducible and defensible to a regulator, not an LLM's say-so. On a fixed, labeled 120-claim synthetic benchmark, 35 of the frauds carry no fake signature at all (recycled-real and metadata-spoofed). A single-signal baseline catches 0 of those 35. The firewall catches all 35, at a 0% false-accusation rate on the clean claims. Those are properties of this constructed set, not an open-world promise, and every claim, photo, and prior in it is synthetic. This runs before damage assessment, not after. It authenticates the evidence Tractable and CCC then assess, it does not replace them. If your claims or SIU team is weighing how to trust claim evidence before acting on it, we would genuinely like to hear how you are approaching it. #InsurTech #ClaimsFraud #SIU #EvidenceIntegrity #AIGovernance