> E.g. imagine somewhere in the code for your chess supercomputer is a genuine bug - some dumb typo, I don't know - which leads it to make the wrong decisions i

> E.g. imagine somewhere in the code for your chess supercomputer is a genuine bug - some dumb typo, I don't know - which leads it to make the wrong decisions in some situations. If the evaluation function executed at the leaf nodes of the search tree contains a bug that massively over-evaluates a position, you might be able to find this by having the chess computer play against a different computer that exploits this bug. Then you'd observe the positions at the end of the principal variation as some other comments mentioned. Even if you can do all that, that still leaves the main problem unanswered which is to explain the correct moves (or rather, moves "assumed to be correct").