Test Before You Deploy: Governing Silent Updates in the Hosted LLM Supply Chain

arXiv·medium signal

Documents a critical operational risk: hosted LLM services (OpenAI, Anthropic, Google) evolve continuously through provider-side updates without explicit version changes, introducing behavioral drift that silently breaks downstream systems. The paper proposes governance frameworks for detecting and managing these silent updates before they reach production. Essential reading for anyone depending on hosted LLM APIs — if you don't test for behavioral regression between API calls, you're flying blind.