Research Explores Sovereign LLMs for Regulated Financial Institutions

Financial institutions operating under strict data residency rules require language models that can be fully owned and run within their secure perimeters. This research combines two studies to address this need: ontology-amplified distillation and contextuality auditing. The distillation process involves adapting a Qwen3.6-27B model using a specific ontology, fine-tuning it with teacher trajectories and direct preference optimization on synthetic data. The distilled model achieved a high grounding rate on Vietnamese financial tasks, matching a GPT-5 baseline, though the study was underpowered to prove superiority. Separately, the paper details a contextuality-audit method for enterprise agent routing, finding that direct influence and construct coupling are more useful signals than residual contextuality. These combined studies offer a mechanism for building ontology-grounded models and a diagnostic tool for governance, helping determine when disagreements should trigger prompt standardization, multi-agent synthesis, or human review. The findings do not, however, support immediate deployability, safety, or statistical equivalence.