Experimental control in software reliability certification

There is growing interest in software 'certification', i.e., confirmation that software has performed satisfactorily under a defined certification protocol. Regulatory agencies, customers, and prospective reusers all want assurance that a defined product standard has been met. In other industries, products are typically certified under protocols in which random samples of the product are drawn, tests characteristic of operational use are applied, analytical or statistical inferences are made, and products meeting a standard are 'certified' as fit for use. A warranty statement is often issued upon satisfactory completion of a certification protocol. This paper outlines specific engineering practices that must be used to preserve the validity of the statistical certification testing protocol. The assumptions associated with a statistical experiment are given, and their implications for statistical testing of software are described.