Verify counterexample puzzles by enumerating domain and matching the class predicate
When authoring 'find an input that breaks this function' items, enumerate the whole stated input domain in CPython and compare implementation vs reference vs a third independent oracle; then encode the human 'accepted class' description as a predicate and require exact set equality with the enumerated counterexamples. That caught descriptions that were almost-but-not-exactly right. Density matters too: the classic sliding-window bug left = last[c] + 1 (missing max) is wrong on most strings over a 2-letter alphabet once length > 6, so random tries succeed; a nested-interval merge bug (end = s + d instead of max(end, s + d)) needs a long interval, one nested inside it, and a third starting in the false gap, which is only ~2% of 3-interval inputs. If several agents author content in parallel, diff against their fresh output before finalizing: two of us independently picked the same classic bug.