`torch_dtype` is deprecated! Use `dtype` instead!
easy: 25/48

== easy: n=48 acc=0.604 cci=44.7 brier=0.5712 ece=0.1767 p50=0.044s p95=0.098s
   extraction             10/12 = 0.833
   fact                   9/12 = 0.750
   intent                 4/12 = 0.333
   tool_selection         6/12 = 0.500
hard: 25/111
hard: 50/111
hard: 75/111
hard: 100/111

== hard: n=111 acc=0.360 cci=3.6 brier=0.7972 ece=0.2586 p50=0.046s p95=0.051s
   adversarial            3/6 = 0.500
   ambiguous              2/7 = 0.286
   judge_hard             8/17 = 0.471
   long_policy            6/19 = 0.316
   multi_hop              7/18 = 0.389
   probability            4/10 = 0.400
   routing_hard           2/5 = 0.400
   temporal_numeric       3/15 = 0.200
   tradeoff               2/6 = 0.333
   trap                   3/8 = 0.375
original: 25/72
original: 50/72

== original: n=72 acc=0.458 cci=21.4 brier=0.7611 ece=0.2717 p50=0.045s p95=0.046s
   adequacy               6/12 = 0.500
   extraction             6/12 = 0.500
   intent                 6/12 = 0.500
   ordinal                6/12 = 0.500
   policy                 5/12 = 0.417
   routing                4/12 = 0.333
wrote /root/alea/jevbench_results.json