Every asset whose bytes the drawn frame depends on, ascending by path.
Ordering is by UTF-16 code unit, never by locale collation, so a Turkish or Swedish host does not fingerprint a different frame than an English one.
Digest of the represented fields at sealing time; subsequent shared-record mutation is not reflected automatically.
Versioned fingerprint protocol.
Who draws the frame.
Renderer configuration that changes what a frame costs.
OptionaltextureHost-decoded upload fact identity. Version two targets include this digest; legacy version one targets omit it and cannot certify those host facts.
The exact renderer, settings, and asset bytes a render report was measured against.
A budget verdict is only evidence while the thing it measured is still the thing that will be drawn. Change the shadow filter, the pixel ratio, or one texture's bytes and the same design costs something else, so a report that outlives its target is not conservative, it is wrong in an unknown direction. The fingerprint makes that detectable: a consumer compares the report's target against the target in front of it and treats a mismatch as stale rather than as a pass.
Identical declared renderer, settings, asset and optional texture facts produce the same fingerprint across hosts. Fixed-position JSON tuples preserve embedded tabs and newlines, and asset ordering uses UTF-16 code units rather than locale collation.
Sealing validates supplied facts rather than acquiring a renderer or reading assets; closure completeness and byte provenance belong to the caller. The returned settings and asset array are new, but renderer and asset records remain shared and mutable, so a later mutation can make the stored digest stale. Target comparison checks the represented fields rather than recomputing or authenticating their digests.