model upgrade testing

LLM upgrades are deceptively risky. Here’s why traditional engineering fails and what eval-first architecture solves.