points
Why not check the logprobs of the output and take action when the prob of the first and second most likely token is too similar? (or below a certain threshold?