I do a lot of numerical programming. When developing programs based on optimization, in particular, a similar robustness-to-error property happens. Your implementation can have bugs, but it's generally hill-climbing, and so often the results generally look OK.
If you really want to verify correct operation, you have to construct hard cases, or compare with another implementation, or look at intermediate state variables, or examine the cost function at very high numerical precision, to detect trouble. Run-of-the-mill inputs will not tickle the bug hard enough to notice.
I do a lot of numerical programming. When developing programs based on optimization, in particular, a similar robustness-to-error property happens. Your implementation can have bugs, but it's generally hill-climbing, and so often the results generally look OK.
If you really want to verify correct operation, you have to construct hard cases, or compare with another implementation, or look at intermediate state variables, or examine the cost function at very high numerical precision, to detect trouble. Run-of-the-mill inputs will not tickle the bug hard enough to notice.