DeepMind and Stanford researchers report that adding a red heart to the same tree-traversal task shifted some models from iteration to recursion. Across 20 runs per condition, Opus 4.7 and GPT-5.4 used iteration without an image and recursion with the heart. Neither implementation is inherently better in every case. Reasoning explanations also failed to reliably reflect the actual visual influence, suggesting that screen-reading coding agents should be tested repeatedly with decorative elements varied. Reproduction across real-world workflows remains unconfirmed.