I am noticing that for our fintech use case, well-written tools on self-hosted models are working just as well as frontier models. But people keep talking about round count and effective token utilization. Suggestions on how I can validate the hypothesis and close the debate for good in our org?