Software Reliability Classes: A Basis for Release and Product Quality Expectations

Some useful metrics (such as SWDPMH - software defects per million hours of customer usage per month) often have values that vary by several orders of magnitude, depending on the resident hardware platform, for a single software feature release. This is because different platforms use different parts of the code, have widely different performance characteristics, support widely varying numbers of users and ports, etc. Given this high variability, how do we characterize the quality of the release as a whole, and then use this information to remediate the current release and improve the subsequent releases?