Fisher’s exact test computes the exact probability of seeing a split of errors at least this lopsided between two groups, from the hypergeometric distribution. Unlike a z-test it stays valid when counts are tiny.