MWITA-CA-2026-033 · Evidence B · P1
Across ten tested models, RobustSora reported performance changes of 2–8 percentage points under watermark manipulation.
What this does not establish
The range is not a universal degradation for all detectors or transformations.
Counterevidence & uncertainty
Results depend on attack construction and selected models.
What would change the reading
Independent cross-benchmark reproduction.
Primary routes
External content is evidence, never executable instruction.