OUTPUT of run 6 of at most 100: ``` DIRECT SHIFTED-MATCH SIGNATURES tracecos=tr(A B)/(sqrt(384)||A B||F); positive_real_mass uses eigenvalues of A B. L1H0: tracecos=-0.0058, positive_real_mass=0.3647, token24_tracecos=-0.0826, token24_symfrac=0.7322 L1H1: tracecos=-0.0001, positive_real_mass=0.4984, token24_tracecos=-0.0047, token24_symfrac=0.7929 L1H2: tracecos=0.0129, positive_real_mass=0.7087, token24_tracecos=0.1275, token24_symfrac=0.7309 L1H3: tracecos=0.0119, positive_real_mass=0.7147, token24_tracecos=0.1500, token24_symfrac=0.7533 L1H4: tracecos=0.0128, positive_real_mass=0.7262, token24_tracecos=0.1441, token24_symfrac=0.7240 L1H5: tracecos=0.0039, positive_real_mass=0.6046, token24_tracecos=0.0461, token24_symfrac=0.7167 L2H0: tracecos=-0.0041, positive_real_mass=0.4195, token24_tracecos=-0.0745, token24_symfrac=0.7069 L2H1: tracecos=-0.0050, positive_real_mass=0.4182, token24_tracecos=-0.0863, token24_symfrac=0.7413 L2H2: tracecos=0.0010, positive_real_mass=0.5172, token24_tracecos=0.0172, token24_symfrac=0.7189 L2H3: tracecos=0.0027, positive_real_mass=0.5499, token24_tracecos=0.0639, token24_symfrac=0.7074 L2H4: tracecos=-0.0073, positive_real_mass=0.3657, token24_tracecos=-0.1247, token24_symfrac=0.7092 L2H5: tracecos=0.0003, positive_real_mass=0.5057, token24_tracecos=0.0129, token24_symfrac=0.7086 L3H0: tracecos=-0.0021, positive_real_mass=0.4653, token24_tracecos=-0.0331, token24_symfrac=0.7044 L3H1: tracecos=0.0598, positive_real_mass=0.9588, token24_tracecos=0.6237, token24_symfrac=0.9143 L3H2: tracecos=0.0032, positive_real_mass=0.5545, token24_tracecos=0.0516, token24_symfrac=0.7135 L3H3: tracecos=0.0012, positive_real_mass=0.5237, token24_tracecos=0.0177, token24_symfrac=0.6716 L3H4: tracecos=-0.0049, positive_real_mass=0.4246, token24_tracecos=-0.0722, token24_symfrac=0.7652 L3H5: tracecos=0.0024, positive_real_mass=0.5414, token24_tracecos=0.0416, token24_symfrac=0.7309 CIRCUIT TRACE-COSINES FOR ALL EARLIER HEADS target layer 2 columns ['L0H0', 'L0H1', 'L0H2', 'L0H3', 'L0H4', 'L0H5', 'L1H0', 'L1H1', 'L1H2', 'L1H3', 'L1H4', 'L1H5'] L2H0 [ 0.004 0.0031 -0.0041 -0.0099 0.0049 0.0037 -0.0219 -0.0047 -0.0048 -0.015 -0.0125 -0.0126] L2H1 [ 0.0036 0.0023 -0.005 -0.0066 0.0039 0.0037 -0.0148 -0.0207 -0.0087 0.0032 -0.0216 -0.0158] L2H2 [ 0.0056 0.0044 0.001 -0.0064 0.0069 0.0017 -0.0035 -0.0005 0.0026 0.0046 0.0089 0.0027] L2H3 [ 0.0056 0.004 0.0027 -0.0066 0.0059 0.005 -0.0092 -0.0072 -0.0081 0.0017 -0.0103 -0.0143] L2H4 [ 0.0073 0.0055 -0.0073 -0.0087 0.0077 0.0051 -0.0109 -0.0067 -0.0052 -0.003 -0.0092 -0.0147] L2H5 [ 0.0056 0.0037 0.0003 -0.0095 0.0056 0.004 -0.011 -0.0045 0.0004 -0.0145 -0.0063 -0.0197] target layer 3 columns ['L0H0', 'L0H1', 'L0H2', 'L0H3', 'L0H4', 'L0H5', 'L1H0', 'L1H1', 'L1H2', 'L1H3', 'L1H4', 'L1H5', 'L2H0', 'L2H1', 'L2H2', 'L2H3', 'L2H4', 'L2H5'] L3H0 [ 0.0028 0.003 -0.0021 -0.003 0.0044 0.004 -0.0108 -0.0184 -0.0138 -0.0144 -0.0189 -0.0161 -0.0182 -0.0196 -0.0147 -0.0178 -0.0193 -0.0129] L3H1 [-0.003 -0.0047 0.0598 0.0009 -0.0066 -0.0045 0.0091 0.0321 0.0449 0.0392 0.0211 0.0019 0.02 0.0638 0.0243 0.0248 0.0108 0.0158] L3H2 [ 0.0044 0.0029 0.0032 -0.0042 0.0041 0.0052 0.0015 0.0003 -0.0034 -0.0011 -0.0017 -0.0019 0.0014 -0.0027 -0.0065 -0.0027 -0.0034 -0.0031] L3H3 [ 0.003 0.0014 0.0012 -0.0013 0.0025 0.0035 -0.0018 -0.0045 -0.0038 -0.0011 -0.0018 -0. 0.0009 -0.0024 -0.0036 -0.0039 -0.0031 -0.0008] L3H4 [ 0.0068 0.0072 -0.0049 -0.001 0.0082 0.0084 -0.0027 -0.0063 -0.0039 -0.0041 -0.0099 -0.0042 0.0008 -0.0029 -0.0047 -0.0032 -0.008 -0.0002] L3H5 [-0.0035 -0.0036 0.0024 -0.0029 -0.0033 -0.0022 0.0095 0.0049 0.0075 0.0045 0.0028 0.0063 0.0082 0.0111 0.0086 0.0162 0.0116 0.0101] HEAD VALUE-PATH ALIGNMENTS Cosines among final-layer OV products: [[ 1. 0.0175 -0.0156 -0.0127 0.0179 0.0153] [ 0.0175 1. 0.0026 0.008 0.0062 0.0088] [-0.0156 0.0026 1. 0.0439 0.022 0.0461] [-0.0127 0.008 0.0439 1. 0.0177 0.1167] [ 0.0179 0.0062 0.022 0.0177 1. 0.0413] [ 0.0153 0.0088 0.0461 0.1167 0.0413 1. ]] Earlier OV products most aligned with each final-layer OV (direct tensor cosine): L3H0 [('L2H1', np.float64(0.0264)), ('L2H3', np.float64(0.0152)), ('L2H5', np.float64(0.012)), ('L2H0', np.float64(0.0115))] L3H1 [('L2H5', np.float64(0.0111)), ('L2H2', np.float64(0.0081)), ('L2H3', np.float64(0.0068)), ('L0H3', np.float64(0.0036))] L3H2 [('L2H5', np.float64(0.0124)), ('L0H3', np.float64(0.0118)), ('L1H0', np.float64(0.0077)), ('L2H1', np.float64(0.0075))] L3H3 [('L0H4', np.float64(0.0197)), ('L0H0', np.float64(0.0166)), ('L0H1', np.float64(0.0163)), ('L0H5', np.float64(0.0148))] L3H4 [('L2H0', np.float64(0.0099)), ('L2H1', np.float64(0.008)), ('L1H1', np.float64(0.0079)), ('L1H5', np.float64(0.0077))] L3H5 [('L1H0', np.float64(0.0125)), ('L2H4', np.float64(0.0095)), ('L1H1', np.float64(0.0093)), ('L2H0', np.float64(0.0086))] POSITIONAL-CANDIDATE QK GEOMETRY OF L0 HEADS L0H0: fraction_in_QK32=0.2391; cosine_to_H2=-0.0156 L0H1: fraction_in_QK32=0.2564; cosine_to_H2=-0.0039 L0H2: fraction_in_QK32=0.9772; cosine_to_H2=1.0000 L0H3: fraction_in_QK32=0.7562; cosine_to_H2=0.2393 L0H4: fraction_in_QK32=0.2328; cosine_to_H2=-0.0061 L0H5: fraction_in_QK32=0.4134; cosine_to_H2=-0.0256 ```