Researchers propose RegLLM, a diagnostic harness for evaluating bounded autonomy in regulated agentic workflows. It instruments six trustworthiness signals and blocks ungrounded answers to improve safety and escalation.