Build a simulation environment that models the recursive collapse of LLM training data. This tool should allow researchers to visualize how synthetic data drift affects model entropy over multiple generations.