What’s a good prediction? Challenges in evaluating an agent’s knowledge