When the requirements for a part are specified, it is based on assumptions that may or may not hold true.
For example, if an issue tends to be all or nothing, then testing a small percentage of a lot should reasonably be expected to catch an issue. So you might specify that 1% of these transistors be tested and so long as that 1% passes the rest are considered good. If let's say there's a process change and lots become more variable, the confidence with which you can say the others are good based on that 1% testing goes down, but you are still testing to the same standard that you were before, which is what the specification calls for.
The issue gets even more thorny when issues are conditional. For example a part might meet the voltage specification, the temperature specification, and the radiation specification individually, but when you put that same part simultaneously in a high voltage, low temperature, and high radiation environment it doesn't perform as well. Or perhaps one component used downstream of a particular other component has an effect. Perhaps the most basic example is oversized but in tolerance shaft meets undersized but in tolerance hole.
>For example a part might meet the voltage specification, the temperature specification, and the radiation specification individually, but when you put that same part simultaneously in a high voltage, low temperature, and high radiation environment it doesn't perform as well.
I am not disagreeing, but at some point humanity should switch to actual probabilistic device models instead of vague datasheets. Imagine every datasheet has a .sample() function and you get a randomized SPICE model as if it came from the manufacturing line, you can draw a 100 and plot the properties. Want to measure the dynamic range of some ADC? A-weighted or not? Instead of specifying values highlighting figures of merit, each such figure of merit corresponds with an explicit SPICE circuit that measures that figure of merit on one and the same generator for random SPICE models with that specific part designation.
If a brand tries to fool its customers by insinuating a desirably high or low value by changing the test method, its immediately clear. A user may specify his own test circuit, or reuse the test circuit from the interactive datasheet from a competitor etc in order to make apples to apples comparisons for different DUT's.
Either the parts were in spec or they weren't. Which is it?