Benchmark Reliability Becomes a Deployment Bottleneck

Epoch AI's reported audit of 15 benchmarks found only four verified, nine flawed, and two unclear, underscoring the fragility of headline model comparisons. The initiative could improve evaluation discipline, but methodology, full reports, corrections, and connections to real-world task performance remain to be established.