Physical randomness generators

It is almost universally assumed that at some level physical processes must be the best potential sources of true randomness. But in practice their record has actually been very poor. It does not help that unlike algorithms physical devices can be affected by their environment, and can also not normally be copied identically. But in almost every case I know where detailed analysis has been done substantial deviations from perfect randomness have been found. This has however typically been attributed to engineering mistakes—or to sampling data too quickly—and not to anything more fundamental that is for example worth describing in publications.