The green agent distributes the images to participating agents and compares their outputs against a predefined ground-truth annotation. Performance is measured based on accuracy and robustness in extracting and interpreting the intended text under visual ambiguity.The green agent evaluates participating agents on their ability to correctly interpret and recognize textual content embedded in visually misleading images. These images are intentionally designed to induce common failure modes in machine perception systems (e.g., ambiguous typography, visual illusions, unconventional layouts), while remaining readily understandable to human readers.