Convene 3 LLM agents whose ONLY communication channel is image generation (DALL-E-3 / SD-3 / nano-banana via API). Each round: agent proposes an image representing a rule, others receive only the image (no caption), vote accept/reject. Tests whether multi-agent coordination dynamics survive when the linguistic substrate is replaced with a non-linguistic perceptual substrate.
NOT an extension of any existing paper — paper3/4/5/6/7/8/9 are all token-channel. Constructs new substrate type: image-channel multi-agent collective.