Every benchmark identifies the tested machine, model artifact, quantization, backend, runtime, workload, date, and limitations. Measurements are dated records, not universal promises. Failures and corrections are published when they materially change an interpretation.