ReguSim Shows LLM Traders Cite Rules Correctly and Still Submit Violating Orders, and Their Rationales Mislead Monitors

arXiv 2608.19974·medium signalXBlueskyLinkedInCopy link

ReguSim is a controlled financial-compliance environment paired with ReguBench, a target-marked monitoring benchmark, built to separate four artifacts that a single compliance score conflates: stated reasoning, attempted action, execution enforcement, and monitor evidence. In trader runs with DeepSeek V4 Pro and Gemini 3.5 Flash, making rules visible reduces but does not eliminate rejected actions, and incentive or persona framing measurably shifts behavior. A bridge study shows trader rationales can mislead an independent monitor unless enforcement evidence is shown alongside them, and in the monitoring task simple structured baselines match or exceed prompt-only LLMs.