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

== easy: n=48 acc=0.583 cci=41.8 brier=0.5109 ece=0.1764 p50=0.033s p95=0.039s
   extraction             6/12 = 0.500
   fact                   8/12 = 0.667
   intent                 6/12 = 0.500
   tool_selection         8/12 = 0.667
hard: 25/111
hard: 50/111
hard: 75/111
hard: 100/111

== hard: n=111 acc=0.306 cci=0.0 brier=0.8489 ece=0.3165 p50=0.033s p95=0.098s
   adversarial            3/6 = 0.500
   ambiguous              2/7 = 0.286
   judge_hard             8/17 = 0.471
   long_policy            6/19 = 0.316
   multi_hop              4/18 = 0.222
   probability            4/10 = 0.400
   routing_hard           2/5 = 0.400
   temporal_numeric       2/15 = 0.133
   tradeoff               2/6 = 0.333
   trap                   1/8 = 0.125
original: 25/72
original: 50/72

== original: n=72 acc=0.444 cci=19.4 brier=0.8043 ece=0.2786 p50=0.032s p95=0.035s
   adequacy               4/12 = 0.333
   extraction             8/12 = 0.667
   intent                 6/12 = 0.500
   ordinal                4/12 = 0.333
   policy                 6/12 = 0.500
   routing                4/12 = 0.333
wrote /root/alea/jevbench_results.json