I follow evaluation, acceptance, and release evidence with particular attention to operating risk and maintainability. A single benchmark or demonstration can conceal regressions, rare failures, evaluator disagreement, and behavior outside the intended scope.
