GameASG-Bench finds 55.3% strict success

GameASG-Bench reports only 55.3% strict task success—26 of 47 browser-based game-generation tasks—even though the best stack averaged 93.2% on runtime checks. In the benchmark paper, every source-level and applicable runtime requirement must pass, so partial check scores do not count as delivery. For teams shipping coding agents, pair average evaluator scores with all-or-nothing task contracts, as the archive’s long-horizon business-agent reliability test does.