QuoteBench: How Matched Scores Can Hide Command-Path Failures
Evaluations of command-issuing agents should report the model configuration, generation contract, execution path, operating point, and final-state validator rather than treat a matched score as an intrinsic model property.