Inside Deep Scan: log-assisted attribution when UID recovery is weak

Normal Scan is designed for speed. It is what teams use most of the time in production.

Deep Scan is for difficult files: heavily compressed uploads, multi-hop reposts, crops, trims, and other transforms that degrade direct UID recovery.

In those cases we run a stronger recovery pass and compare recovered signal traits and media characteristics against internal encode/decode records.

This does not invent certainty where there is none. It produces an attribution confidence score with an evidence package so reviewers can make policy decisions with context.

For trust and safety teams, this is critical: a degraded file can still be assessed against known generation records instead of being treated as an untraceable dead end.

Roadmap note: we are also training models to estimate AI-origin likelihood in cases where no watermark is present.