Mechanical Enforcement Improves LLM Governance in Finance
A recent preprint on arXiv (2605.14744) presents five metrics designed to measure policy adherence at the decision rationale level for large language models utilized in regulated financial processes. The research indicates that governance relying solely on text is hindered by a principal-agent issue, leading to 27% of deferrals lacking relevant decision information. By implementing mechanical enforcement through four primitives that function outside the model's interpretive framework, this deficiency is reduced by 73%, while the information content of deferrals more than doubles and task accuracy improves from MCC 0.43 to 0.88. This enhancement is credited to the architectural separation of governance from the LLM.
Key facts
- arXiv:2605.14744 introduces five governance metrics for LLM compliance in finance.
- Text-only governance leads to 27% of deferrals with no decision-relevant information.
- Mechanical enforcement reduces non-informative deferrals by 73%.
- Task accuracy improves from MCC 0.43 to 0.88 with mechanical enforcement.
- Four primitives operate outside the model's interpretive loop.
- The study addresses principal-agent failure in LLM governance.
- Governance metrics quantify compliance at the decision rationale level.
- Architectural separation of governance from LLM drives improvement.
Entities
Institutions
- arXiv