A preprint reports that EchoCoT recovered hidden AI reasoning through tool calls, with strong results on accessible models but indirect evidence for proprietary ones