I never said it was an honest measure, but it is the measure used by the hard drive industry. It is also very similar to quality measures in many other industries. You take a reasonable sample, run a test for a certain amount of time and infer/extrapolate reliability figures based on the test. On the back-end, you redesign your process to eliminate as many variations as possible on the manufacturing side.
The alternative would be to really test the disks for 3-5 years in a lab before releasing them on the market, or really testing a large amount of drive trains for 10 years and so many miles before releasing to market. The first company to really do that will go bankrupt before shipping the next generation of devices.