claude-opus-5-5 train problems 65 mean lift -0.069 [-0.120, -0.022] neg 25 pos 12 | by family: attention -0.17 (n5), conv -0.11 (n14), conv_transpose -0.06 (n10), matmul -0.06 (n22), network +0.02 (n9), norm -0.02 (n1), rnn -0.08 (n4) claude-opus-5-5 gate problems 27 mean lift -0.185 [-0.342, -0.038] neg 14 pos 2 | by family: attention -0.16 (n1), conv -0.40 (n7), conv_transpose -0.24 (n4), matmul -0.10 (n9), network -0.03 (n4), rnn -0.04 (n2) claude-opus-5-5 train+gate problems 92 mean lift -0.103 [-0.163, -0.048] neg 39 pos 14 | by family: attention -0.17 (n6), conv -0.20 (n21), conv_transpose -0.11 (n14), matmul -0.07 (n31), network +0.00 (n13), norm -0.02 (n1), rnn -0.07 (n6) claude-opus-4-8 train problems 55 mean lift +0.204 [+0.121, +0.285] neg 6 pos 30 | by family: attention +0.10 (n5), conv +0.34 (n11), conv_transpose +0.30 (n8), matmul +0.16 (n18), network +0.16 (n9), norm -0.65 (n1), rnn +0.27 (n3) claude-opus-4-8 gate problems 22 mean lift +0.387 [+0.212, +0.564] neg 1 pos 14 | by family: attention +0.59 (n1), conv +0.62 (n5), conv_transpose +0.72 (n4), matmul +0.07 (n8), network +0.09 (n2), rnn +0.61 (n2) claude-opus-4-8 train+gate problems 77 mean lift +0.256 [+0.178, +0.337] neg 7 pos 44 | by family: attention +0.18 (n6), conv +0.43 (n16), conv_transpose +0.44 (n12), matmul +0.14 (n26), network +0.14 (n11), norm -0.65 (n1), rnn +0.41 (n5)