Qwen3.6-27B Distilled to Match GPT-5 on Financial Tasks in Sovereignty Study

A recent arXiv preprint (2607.11948) by Thanh Luong Tuan combines two related studies into one proof-of-mechanism and negative-results article. First, the authors use ontology-amplified distillation to adapt a Qwen3.6-27B student to the Foundation AgenticOS ontology. The training pipeline includes supervised fine-tuning on frontier-teacher trajectories and ontology-grounded direct preference optimization (DPO), all performed locally on a single Apple M5 Max with just 47 synthetic, English-language, cross-domain preference pairs. On 40 held-out Vietnamese financial-domain tasks, the distilled student grounds 36 out of 40 tasks (grounded rate 0.90; mean ontology term-coverage r_onto = 0.95), exactly matching the GPT-5 frontier baseline. However, due to the small sample size, the paired-difference 95% confidence interval spans ±4 tasks, meaning the study is underpowered to establish statistical equivalence or superiority.

The second part of the paper consolidates a contextuality-audit method for enterprise-agent routing. In a separate negative-results pilot, the corrected canonical Contextuality-by-Default degree was zero for all Phase 1.3 groups in both the local-Qwen run and a Gemma replication check. This indicates that the useful signal comes from direct influence and construct coupling, not from surviving residual contextuality. Together, the studies pair an ontology-grounded model-building mechanism with a governance diagnostic for deciding when apparent disagreement should trigger prompt standardization, multi-agent synthesis, or human review. The authors are clear that the evidence supports neither deployability, safety, superiority, statistical equivalence, nor a contextuality-positive routing rule.