Software fault-freeness and reliability predictions

Many software development practices aim at ensuring that software is correct, or fault-free. In safety critical applications, requirements are in terms of probabilities of certain behaviours, e.g. as associated to the Safety Integrity Levels of IEC 61508. The two forms of reasoning - about evidence of correctness and about probabilities of certain failures -are rarely brought together explicitly. The desirability of using claims of correctness has been argued by many authors, but not been taken up in practice. We address how to combine evidence concerning probability of failure together with evidence pertaining to likelihood of fault-freeness, in a Bayesian framework. We present novel results to make this approach practical, by guaranteeing reliability predictions that are conservative (err on the side of pessimism), despite the difficulty of stating prior probability distributions for reliability parameters. This approach seems suitable for practical application to assessment of certain classes of safety critical systems