arXiv paper reports ontology-amplified distillation for enterprise LLMs but finds no evidence of superiority or deployability

A new arXiv paper combines a reduced-power proof-of-mechanism study of ontology-amplified distillation with a negative-results pilot on contextuality auditing for enterprise-agent routing. The distillation study adapts a Qwen3.6-27B student to the Foundation AgenticOS ontology using supervised fine-tuning on frontier-teacher trajectories and ontology-grounded direct preference optimization (DPO), trained locally on a single Apple M5 Max workstation using 47 synthetic, English-language, cross-domain preference pairs.

On 40 held-out Vietnamese financial-domain tasks, the distilled student achieved a grounded rate of 0.90 (36 of 40 tasks grounded) and a mean ontology term-coverage r_onto of 0.95 (metric floored at 0.50), matching the performance of a GPT-5 frontier baseline, which also grounded 36 of 40 tasks. The authors note the study is underpowered to establish statistical equivalence: the paired-difference 95% confidence interval spans +/-4 tasks, and the run did not test or show the pre-registered amplification prediction that the student should exceed the frontier.

The companion negative-results pilot consolidates a contextuality-audit method for enterprise-agent routing. In a separate analysis, the corrected canonical Contextuality-by-Default degree was zero for all Phase 1.3 groups in both the local-Qwen run and an explicitly labeled Gemma replication check. The authors interpret the useful signal as direct influence and construct coupling rather than surviving residual contextuality.

Together, the studies pair an ontology-grounded model-building mechanism with a governance diagnostic aimed at deciding when apparent disagreement should trigger prompt standardization, multi-agent synthesis, or human review. The authors explicitly state the evidence supports neither deployability, safety, superiority, statistical equivalence, nor a contextuality-positive routing rule.