VIDRAFT has launched ai-world (CIVOS), a live multi-agent research platform designed to differentiate between genuine emergent discovery and mere training-data recall in large-scale AI simulations. Operating with tens of thousands of agents, the platform employs a rigorous experimental framework using controlled ablation of agent knowledge. The methodology utilizes three primary conditions: a baseline with full pre-trained knowledge, a knowledge-removed condition, and a knowledge-corrupted condition. Research findings indicate that when prior knowledge is removed, the formation of civilization-like structures degrades significantly. Furthermore, the knowledge-corrupted condition performs worse than the zero-knowledge condition, suggesting that incorrect prior knowledge actively misleads agent behavior rather than fostering independent exploration. VIDRAFT notes that full-knowledge runs show high consistency across different random seeds, which they interpret as evidence of retrieval rather than novel exploration. While specific numeric effect sizes for the ai-world experiments were not provided, VIDRAFT highlighted their broader technical standing, including 1.6 million cumulative Hugging Face downloads and a 90.9% GPQA score. The platform is accessible on Hugging Face for public observation, allowing researchers to scrutinize the methodology and propose counterexamples to the hypothesis that observed multi-agent patterns are often just trained recitations of human history.