Forensic pipeline
How a scan works, from upload to report.
Sealed Rose inspects digital video, images, and audio for generative AI artifacts, deepfakes, face swaps, and biometric tampering. Here is how our four-stage verification process operates from ingest to final forensic verdict.
Ingest & ephemeral buffer
Video, image, or audio input is ingested directly into volatile memory. SHA-256 cryptographic fingerprints are calculated immediately. Files are deleted right after your scan, except images you send to search engines (kept up to 1 hour for pivot lookups) and confirmed CSAM evidence (retained up to 1 year for mandatory NCMEC compliance).
Multi-frame neural analysis
Video clips are deconstructed frame-by-frame across 10 to 45 sampled frames, combining foundation classification by our detection partners with our own temporal lattice and optical-flow checks (Wan 2.1, Kling, Sora, Runway Gen-3).
Acoustic frequency & voice analysis
Audio tracks undergo spectral analysis and acoustic classification with our detection partners to detect synthetic speech and voice cloning (ElevenLabs, Cartesia, and open-source models).
Probabilistic scan report & deletion
Detection partner classifications and frame signals are aggregated into a transparent scan report with probabilistic confidence scores. Media buffers are permanently deleted right after your scan, except search pivots (kept up to 1 hour) and confirmed CSAM evidence (retained up to 1 year for mandatory NCMEC compliance).
SHA-256: 8f2c...41e9 · Multi-frame neural slicing
Wan 2.1 / Kling high-frequency lattice artifact match
Spectrogram indicates neural voice vocoder resynthesis
Encoding headers indicate synthetic re-rendering
~5–10s
End-to-end processing (1.8–3.4s inference)
Deleted after scan
Deleted right after your scan; pivots kept up to 1h, CSAM 1y
Multi-model
Wan 2.1, Kling, Sora, ElevenLabs & Flux
10–45 frames
Multi-frame video decomposition