A Framework for Vertical Scaling of Domain-Specific Small Language Models for High-Stakes Financial Compliance Using Programmatic Regulatory Instruction Generation

This paper proposes a deterministic framework for vertically scaling domain-specific Small Language Models (SLMs) for financial compliance using programmatic regulatory instruction generation. Instead of relying on AI-generated synthetic datasets, the framework transforms structured financial regulations into reproducible instruction-response pairs through deterministic software logic. Combined with QLoRA-based parameter-efficient fine-tuning, the proposed methodology aims to improve auditability, regulatory consistency, computational efficiency, and enterprise deployment of SLMs in high-stakes financial compliance environments.