Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

Jevstiller is a local machine learning model that learns from an API-based model called Jev to provide fast, low-latency predictions (15ms vs 300ms) while maintaining a configurable agreement guarantee with the original model. The key innovation is a disagreement bound contract that ensures the system agrees with Jev's answers at a specified rate (e.g., 98%) across all requests, which the authors demonstrate outperforms naive confidence threshold approaches that fail to meet the budget guarantees roughly half the time.