marked 17 contention-OOM protected evals as infra === claude-opus-5-5: 317 rows, 37 problems === arm n infra l1p corr tok_out cuda tokK turns sub@1 tok1 late fg_target 27 2 1.186 1.00 9957 0.48 0.04 2.60 0.48 5787 nan frontier_gated 39 3 1.135 1.00 10074 0.47 0.14 2.53 0.61 5691 nan late_show_all 36 3 1.311 1.00 12676 0.91 0.03 2.61 0.45 7081 1.00 lessons_only 34 3 1.282 1.00 10378 0.65 0.10 2.39 0.61 6673 nan no_approach 38 1 1.181 1.00 11164 0.57 0.16 2.41 0.70 6674 nan no_numbers 26 1 1.027 1.00 8779 0.16 0.24 2.36 0.72 5076 nan none 37 0 1.179 1.00 13393 0.89 0.11 2.54 0.59 7665 nan reader_relative 38 1 1.124 1.00 9310 0.35 0.19 2.46 0.62 5517 nan show_all 42 3 1.014 1.00 7082 0.15 0.21 2.31 0.77 4411 nan show_all-none l1p -0.272 [-0.466, -0.081] p=0.0060 n_prob=25 | infra0 -0.327 | refnorm -0.253 [-0.422, -0.078] | tok -4609 | cuda -0.74 | tokK +0.04 lessons_only-none l1p +0.001 [-0.066, +0.073] p=0.9950 n_prob=20 | infra0 -0.072 | refnorm +0.026 [-0.046, +0.101] | tok -802 | cuda -0.28 | tokK +0.03 no_approach-none l1p -0.001 [-0.119, +0.115] p=0.9880 n_prob=21 | infra0 -0.057 | refnorm -0.000 [-0.123, +0.120] | tok -1585 | cuda -0.33 | tokK +0.05 show_all-no_approach l1p -0.244 [-0.450, -0.077] p=0.0000 n_prob=23 | infra0 -0.209 | refnorm -0.225 [-0.435, -0.044] | tok -2928 | cuda -0.41 | tokK +0.00 no_approach-lessons_only l1p -0.122 [-0.271, +0.011] p=0.0790 n_prob=19 | infra0 -0.100 | refnorm -0.137 [-0.278, -0.016] | tok +300 | cuda -0.16 | tokK +0.11 show_all-lessons_only l1p -0.276 [-0.518, -0.002] p=0.0490 n_prob=21 | infra0 -0.282 | refnorm -0.280 [-0.519, -0.007] | tok -3478 | cuda -0.45 | tokK +0.05 no_numbers-show_all l1p +0.002 [-0.112, +0.136] p=0.9510 n_prob=17 | infra0 +0.101 | refnorm -0.000 [-0.111, +0.133] | tok +1047 | cuda +0.03 | tokK +0.03 no_numbers-none l1p -0.193 [-0.374, -0.024] p=0.0200 n_prob=18 | infra0 -0.193 | refnorm -0.202 [-0.384, -0.037] | tok -4856 | cuda -0.67 | tokK +0.14 reader_relative-show_all l1p +0.106 [+0.008, +0.211] p=0.0280 n_prob=24 | infra0 +0.154 | refnorm +0.094 [-0.007, +0.201] | tok +1375 | cuda +0.15 | tokK +0.00 reader_relative-none l1p -0.138 [-0.275, +0.005] p=0.0640 n_prob=21 | infra0 -0.164 | refnorm -0.151 [-0.282, -0.006] | tok -2813 | cuda -0.55 | tokK +0.05 frontier_gated-none l1p -0.118 [-0.306, +0.045] p=0.1840 n_prob=22 | infra0 -0.201 | refnorm -0.112 [-0.286, +0.044] | tok -4256 | cuda -0.41 | tokK +0.00 frontier_gated-show_all l1p +0.103 [-0.068, +0.282] p=0.2480 n_prob=25 | infra0 +0.089 | refnorm +0.089 [-0.080, +0.270] | tok +1772 | cuda +0.30 | tokK -0.06 fg_target-none l1p -0.087 [-0.312, +0.122] p=0.4320 n_prob=19 | infra0 -0.131 | refnorm -0.095 [-0.319, +0.115] | tok -1934 | cuda -0.42 | tokK -0.03 fg_target-show_all l1p +0.157 [+0.007, +0.300] p=0.0420 n_prob=20 | infra0 +0.117 | refnorm +0.152 [+0.004, +0.291] | tok +3514 | cuda +0.33 | tokK -0.07 fg_target-frontier_gated l1p +0.037 [-0.131, +0.212] p=0.6390 n_prob=18 | infra0 +0.034 | refnorm +0.028 [-0.136, +0.203] | tok +331 | cuda -0.11 | tokK -0.11 late_show_all-none l1p +0.106 [+0.008, +0.210] p=0.0330 n_prob=20 | infra0 +0.018 | refnorm +0.133 [+0.037, +0.238] | tok +1325 | cuda +0.03 | tokK -0.05 late_show_all-show_all l1p +0.333 [+0.125, +0.519] p=0.0020 n_prob=24 | infra0 +0.287 | refnorm +0.333 [+0.124, +0.519] | tok +5534 | cuda +0.75 | tokK -0.10