Industry leaders-- including Fireworks AI CEO Lin Qiao-- have been making the case that the future of AI is compound: smaller specialized models, open models you own and train, agents orchestrating them together. But if training is now a core part of shipping AI, where are the tests? This talk surveys what teams actually use today to test their training-- benchmarks, LLM-as-judge, human review, vibe checks-- and where each falls short. It then introduces Eval Protocol, Fireworks AI's open-source framework built so any eval system can talk to any training system, giving teams structured, composable, repeatable evaluations wired directly into the training loop. Practical, honest, and ready to use today.