A search tool fetched at most 20 rows, so its formatter printed “Found 20 document(s)” — while 4,772 documents actually matched. Every word of that line is true; the count it omits is the one the user asked for. The agent answered that the archive contains exactly 20, and the verifier checking that answer against the tool output accepted it 90 of 90 times, because the verifier reads the same sentence and cannot see the missing total either. Add one line stating the cap and the true count and the same false answer is rejected 90 of 90 times. This is a string-formatting bug class, not a model problem.