🤖 AI Summary
This work addresses the critical issue that large language models in financial compliance often produce decisions that appear compliant on the surface yet violate regulations in substance, while existing evaluation frameworks neglect governance constraints on the reasoning behind such decisions. To tackle this, the authors propose a mechanistic governance mechanism that decouples governance from task execution through four external enforcement primitives, ensuring auditability and genuine compliance of the model’s rationales. The study reveals, for the first time, a decoupling between governance quality and task accuracy, demonstrating that high accuracy does not guarantee effective governance. Experiments show that the proposed approach reduces uninformative delayed decisions by 73%, more than doubles the information content of delayed responses, and improves the Matthews Correlation Coefficient (MCC) from 0.43 to 0.88, with causal ablation studies confirming the necessity of each primitive.
📝 Abstract
Large language models in regulated financial workflows are governed by natural-language policies that the same model interprets, creating a principal--agent failure: outputs can appear compliant without being compliant. Existing evaluation measures task accuracy but not whether governance constrains behaviour at the decision rationale level -- where regulated decisions must be auditable. We introduce five governance metrics that quantify policy compliance at the rationale level and apply them in a synthetic banking domain to compare text-only governance against mechanical enforcement: four primitives operating outside the model's interpretive loop. Under text-only governance, 27% of deferrals carry no decision-relevant information. Mechanical enforcement reduces this rate by 73%, more than doubles deferral information content, and raises task accuracy from MCC~$0.43$ to $0.88$. The improvement is driven by architectural separation: LLM-generated rationales under mechanical enforcement show comparable CDL to text-only governance -- the gain comes from removing clear-cut decisions from the model's control. A causal ablation confirms that each primitive is individually necessary. Our central finding is a governance-task decoupling: under structural stress, text-only governance degrades on both dimensions simultaneously, whereas mechanical enforcement preserves governance quality even as task performance drops. This implies that governance and task evaluation are distinct axes: accuracy is not a sufficient proxy for governance in regulated AI systems.