[1] Facial-recognition systems are often tested by comparing images under controlled benchmark conditions. [2] Vendors report aggregate rates for matching known identities or rejecting nonmatches. [3] Civil-rights researchers emphasize differences across demographic groups and image conditions. [4] Deployment settings also vary in camera quality, database size, and the consequences of an error. [5] Aggregate accuracy is an inadequate basis for deployment when it conceals who bears errors and how a mistaken match will be used. [6] Evaluation must connect subgroup performance and operating conditions to the decision the system is allowed to influence.