arm | metric | median | vs base | pairs faster (higher) | rounds
default | batch:tok_s | 28.69 | 1.000 | 0/1 | 1
bg2 | batch:tok_s | 30.26 | 1.055 | 1/1 | 1
default2 | batch:tok_s | 28.84 | 1.005 | 1/1 | 1
default | batch:steady | 32.34 | 1.000 | 0/1 | 1
bg2 | batch:steady | 37.38 | 1.156 | 1/1 | 1
default2 | batch:steady | 32.51 | 1.005 | 1/1 | 1
default | batch:ttft | 0.32 | 1.000 | 0/1 | 1
bg2 | batch:ttft | 0.49 | 1.550 | 0/1 | 1
default2 | batch:ttft | 0.31 | 0.995 | 1/1 | 1

arm | window | mean util per card % (median over runs) | mean power W per card | idle% (mean over cards)
default | batch | gpu0 60, gpu1 42 | 79, 84 | 49
bg2 | batch | gpu0 92, gpu1 85 | 91, 100 | 11
default2 | batch | gpu0 62, gpu1 38 | 80, 84 | 50

answer hashes per window kind (distinct values over all runs of all arms; 1 = byte-identical)
batch: 2 distinct -> c5c805ed94773556:default,default2; 8f596d32d5c6c1ae:bg2
default: expert-cache slots ['2246']
bg2: expert-cache slots ['2246']
default2: expert-cache slots ['2246']
