From capture to a LENS-VERIFIED product.
Bound at the source.
Sensor identity, time, orbit state and exposure are bound to the raw frame's digest the moment it is captured.
Each tile gets its own identity.
Frames are cut into tiles; each tile has a digest and points to its parent frame, so a later product can cite exactly the tiles it used.
Every step is a receipt.
Calibration, orthorectification and classification each record their inputs, their version and their outputs. Swapping a model version shows up.
One leg missing means rejected.
Each receipt carries Ed25519 and ML-DSA-65 signatures over identical canonical bytes. If one leg is missing, the record is rejected, never quietly downgraded (COSRYX-HYBRID-02, 37/37).
Earned, not assumed.
A derived product is marked LENS-VERIFIED only when every receipt in its chain verifies on both legs.
The answer shows its sources.
When the AI ranks a site or flags a hazard, the result cites the tiles and steps behind it. A reviewer can rebuild it.