Benchmark
Computer-use tasks driven through a real desktop — scored as a partial score on the offline subset.