Ontology-Amplified Distillation and Contextuality Auditing for Sovereign Enterprise Language Models: A Combined Proof-of-Mechanism and Negative-Results Method Study

Short summary

This combined proof-of-mechanism and negative-results study tests ontology-amplified distillation for sovereign enterprise LLMs, distilling a Qwen3.6-27B student from frontier-teacher trajectories on a single Apple M5 Max. The distilled student grounds 36 of 40 held-out Vietnamese financial tasks, matching GPT-5's baseline, but the study is underpowered to claim statistical equivalence or superiority. A companion contextuality-audit method for enterprise-agent routing finds zero contextuality signal, with direct influence and construct coupling being the useful diagnostics instead.

• •Qwen3.6-27B distilled via ontology-grounded DPO matches GPT-5 on 40 Vietnamese financial tasks (36/40 each)
• •Study is underpowered: 95% CI spans ±4 tasks, no equivalence or superiority established
• •Contextuality-audit method finds zero contextuality signal; direct influence and construct coupling are the useful diagnostics

Generated with AI, which can make mistakes.