ReguSim and ReguBench separate four artifacts in financial compliance: stated reasoning, attempted action, execution enforcement, and monitor evidence. In trader runs with DeepSeek V4 Pro and Gemini 3.5 Flash, showing rules reduced but did not eliminate rejected orders, while incentive and persona framing shifted behavior.
A bridge study found an independent monitor can be misled by a trader's confident but wrong rationale unless execution evidence is shown. For monitoring, simple structured baselines matched or exceeded prompt-only LLMs. The work argues compliance evaluation should be an audit of rule-grounded actions and evidence use, not a single compliance score.