=== GLOBAL DIFFERENCE ===
Present parameters: 100326016
Exactly changed: 79357570 fraction: 0.7909969234699801
||A||, ||B||, ||B-A||: [22667.0786 22667.0786     2.45  ]
Symmetric relative difference: 0.00010808547312563405

=== LAYER/BLOCK DIFFERENCES ===
layer part        changed%     ||B-A||      rel   delta_energy_share%
   -1 norm         0.0000           0         0             0.00000
    0 attention   74.9499      0.2893 0.0051027             1.39436
    0 mlp         76.7904     0.46838 0.0069013             3.65492
    0 norm         3.7891   0.0045254 1.1036e-05             0.00034
    1 attention   78.2573     0.29007 0.0045523             1.40182
    1 mlp         78.7807      0.4771 0.0072959             3.79217
    1 norm         2.3438   0.0033859 3.1991e-06             0.00019
    2 attention   76.9977     0.31022 0.0052934             1.60330
    2 mlp         78.3534     0.47739 0.0072352             3.79676
    2 norm         1.9922   0.0032846 3.042e-06             0.00018
    3 attention   80.2027     0.30423 0.0063696             1.54193
    3 mlp         79.3560     0.46854 0.0071807             3.65740
    3 norm         1.2891   0.0024661 1.7151e-06             0.00010
    4 attention   80.0727     0.30729 0.0051298             1.57319
    4 mlp         80.7190     0.48121 0.0074273             3.85777
    4 norm         1.1328   0.0026854 1.7542e-06             0.00012
    5 attention   80.3506     0.32017 0.0042941             1.70779
    5 mlp         78.2597     0.48338 0.0071107             3.89264
    5 norm         0.8984   0.0019274 1.5383e-06             0.00006
    6 attention   79.3829     0.32888 0.0061221             1.80194
    6 mlp         79.6054     0.47911 0.0073197             3.82419
    6 norm         0.9766    0.002288 1.8473e-06             0.00009
    7 attention   78.0986     0.31143  0.005449             1.61579
    7 mlp         79.0159     0.47661 0.0072155             3.78443
    7 norm         0.7031   0.0028445 2.3762e-06             0.00013
    8 attention   79.1698      0.3151 0.0055828             1.65416
    8 mlp         80.2335     0.47854 0.0073606             3.81522
    8 norm         1.0547   0.0023546 1.7033e-06             0.00009
    9 attention   78.5027     0.32338 0.0055512             1.74220
    9 mlp         78.8552     0.48008 0.0071931             3.83978
    9 norm         1.0938   0.0022962 1.575e-06             0.00009
   10 attention   77.9812     0.33928 0.0047334             1.91769
   10 mlp         80.2724       0.481 0.0073749             3.85452
   10 norm         0.8203   0.0015124 6.1744e-07             0.00004
   11 attention   79.0360     0.32509  0.003852             1.76069
   11 mlp         79.1145     0.48121 0.0072359             3.85788
   11 norm         0.7812   0.0020267 3.392e-07             0.00007
   12 attention   80.2586     0.33662  0.005574             1.88778
   12 mlp         79.3122      0.4937 0.0073467             4.06067
   12 norm         1.0938    0.002649 3.6137e-07             0.00012
   13 attention   80.1172     0.32046 0.0058845             1.71091
   13 mlp         80.0255     0.50013 0.0076652             4.16712
   13 norm         2.6953   0.0035743 5.2023e-07             0.00021
   14 attention   78.3833     0.31198 0.0047359             1.62150
   14 mlp         80.5829     0.50954 0.0078405             4.32538
   14 norm         3.9062   0.0043347 5.513e-07             0.00031
   15 attention   77.9756     0.33227 0.0052633             1.83935
   15 mlp         79.9643     0.50933 0.0077399             4.32192
   15 norm         1.2891    0.003179 3.2794e-07             0.00017
   16 attention   79.9434     0.31076 0.0056186             1.60891
   16 mlp         78.9666     0.49128  0.007422             4.02102
   16 norm         0.7031   0.0014952 1.5617e-07             0.00004
   17 attention   78.4393     0.29417 0.0044515             1.44169
   17 mlp         78.8741     0.46799 0.0071387             3.64877
   17 norm         0.9766   0.0022775 2.2093e-07             0.00009

=== TOP 18 TENSORS BY ABSOLUTE DIFFERENCE ===
model.layers.14.mlp.gate_proj.weight: delta=0.302144, rel=0.00676324, changed=76.753%, norms=44.6744/44.6746
model.layers.15.mlp.gate_proj.weight: delta=0.299849, rel=0.00670883, changed=76.481%, norms=44.6947/44.6945
model.layers.15.mlp.up_proj.weight: delta=0.29968, rel=0.00670756, changed=74.819%, norms=44.678/44.6779
model.layers.14.mlp.up_proj.weight: delta=0.297807, rel=0.00667512, changed=74.583%, norms=44.6143/44.6147
model.layers.13.mlp.gate_proj.weight: delta=0.297223, rel=0.00665486, changed=76.229%, norms=44.6624/44.6628
model.layers.12.mlp.gate_proj.weight: delta=0.295251, rel=0.006594, changed=77.129%, norms=44.7756/44.7757
model.layers.13.mlp.up_proj.weight: delta=0.294863, rel=0.00661386, changed=74.485%, norms=44.5824/44.5828
model.layers.11.mlp.gate_proj.weight: delta=0.293139, rel=0.00654692, changed=76.932%, norms=44.775/44.7751
model.layers.6.mlp.gate_proj.weight: delta=0.2922, rel=0.00653479, changed=76.493%, norms=44.7144/44.7146
model.layers.5.mlp.gate_proj.weight: delta=0.291332, rel=0.00650045, changed=76.491%, norms=44.8173/44.817
model.layers.12.mlp.up_proj.weight: delta=0.29062, rel=0.00650123, changed=75.114%, norms=44.7023/44.7023
model.layers.16.mlp.up_proj.weight: delta=0.290149, rel=0.00648203, changed=73.998%, norms=44.7616/44.7623
model.layers.16.mlp.gate_proj.weight: delta=0.290139, rel=0.00650715, changed=75.716%, norms=44.5876/44.5879
model.layers.8.mlp.gate_proj.weight: delta=0.289307, rel=0.00644047, changed=76.897%, norms=44.9205/44.9198
model.layers.10.mlp.gate_proj.weight: delta=0.288408, rel=0.00643131, changed=76.757%, norms=44.8444/44.8444
model.layers.4.mlp.gate_proj.weight: delta=0.287782, rel=0.00642344, changed=76.084%, norms=44.802/44.8017
model.layers.9.mlp.gate_proj.weight: delta=0.287289, rel=0.00640056, changed=76.806%, norms=44.8851/44.8849
model.layers.7.mlp.gate_proj.weight: delta=0.286724, rel=0.00640105, changed=76.211%, norms=44.7933/44.7934

=== TOP 12 TENSORS BY RELATIVE DIFFERENCE ===
model.layers.4.mlp.down_proj.weight: rel=0.0187882, delta=0.261557, norms=13.9209/13.9218
model.layers.14.mlp.down_proj.weight: rel=0.0183266, delta=0.282217, norms=15.3993/15.3994
model.layers.8.mlp.down_proj.weight: rel=0.017131, delta=0.254188, norms=14.838/14.8378
model.layers.1.mlp.down_proj.weight: rel=0.0169488, delta=0.263844, norms=15.5669/15.5673
model.layers.10.mlp.down_proj.weight: rel=0.0167458, delta=0.260869, norms=15.5783/15.578
model.layers.13.mlp.down_proj.weight: rel=0.0165041, delta=0.273572, norms=16.5759/16.5761
model.layers.3.mlp.down_proj.weight: rel=0.0162541, delta=0.258182, norms=15.8844/15.8839
model.layers.15.mlp.down_proj.weight: rel=0.0153843, delta=0.282315, norms=18.3509/18.3508
model.layers.6.mlp.down_proj.weight: rel=0.0149201, delta=0.255254, norms=17.1081/17.108
model.layers.2.mlp.down_proj.weight: rel=0.0140623, delta=0.260828, norms=18.5478/18.5483
model.layers.16.mlp.down_proj.weight: rel=0.0136839, delta=0.270168, norms=19.7433/19.7438
model.layers.17.mlp.down_proj.weight: rel=0.0136275, delta=0.248897, norms=18.2643/18.2642

Identical tensors: 32
model.layers.0.input_layernorm.weight
model.layers.0.pre_feedforward_layernorm.weight
model.layers.1.input_layernorm.weight
model.layers.1.pre_feedforward_layernorm.weight
model.layers.1.self_attn.k_norm.weight
model.layers.1.self_attn.q_norm.weight
model.layers.10.post_feedforward_layernorm.weight
model.layers.11.post_attention_layernorm.weight
model.layers.11.post_feedforward_layernorm.weight
model.layers.12.post_attention_layernorm.weight
model.layers.12.post_feedforward_layernorm.weight
model.layers.13.post_attention_layernorm.weight
model.layers.13.post_feedforward_layernorm.weight
model.layers.14.post_attention_layernorm.weight
model.layers.14.post_feedforward_layernorm.weight
model.layers.15.input_layernorm.weight
model.layers.15.post_attention_layernorm.weight
model.layers.15.post_feedforward_layernorm.weight
model.layers.16.input_layernorm.weight
model.layers.16.post_attention_layernorm.weight

=== RAW ATTENTION OPERATOR DIFFERENCES ===
QK = Q_h.T @ K; OV = O_h @ V.
No RoPE, norm diagonals, state normalization, softmax, or inputs.
All four heads share K and V; their differences are not independent.
Per-layer relative differences, heads 0,1,2,3:
L00 QK: 0.00513 0.00679 0.00548 0.00677  OV: 0.00765 0.01106 0.01004 0.00739
L01 QK: 0.00737 0.00834 0.00660 0.00817  OV: 0.01292 0.01209 0.01190 0.01054
L02 QK: 0.00946 0.00949 0.00991 0.01028  OV: 0.01014 0.00995 0.00900 0.01296
L03 QK: 0.01030 0.01355 0.01231 0.00947  OV: 0.01153 0.01406 0.01264 0.01884
L04 QK: 0.01155 0.01274 0.01160 0.01300  OV: 0.01622 0.01768 0.01425 0.01073
L05 QK: 0.01155 0.01000 0.01021 0.01131  OV: 0.00880 0.01144 0.01261 0.01328
L06 QK: 0.00945 0.00965 0.00909 0.00938  OV: 0.01113 0.00922 0.01211 0.01089
L07 QK: 0.00808 0.00897 0.00836 0.00843  OV: 0.01192 0.00873 0.01023 0.01365
L08 QK: 0.01257 0.01188 0.01407 0.01261  OV: 0.01120 0.01260 0.01415 0.01080
L09 QK: 0.01050 0.00994 0.00980 0.00917  OV: 0.01173 0.00866 0.01116 0.01122
L10 QK: 0.01209 0.01059 0.00925 0.01073  OV: 0.01073 0.01035 0.01243 0.00826
L11 QK: 0.01360 0.01411 0.00883 0.01125  OV: 0.01164 0.01277 0.00915 0.00828
L12 QK: 0.00873 0.00887 0.00940 0.00890  OV: 0.01186 0.01105 0.00905 0.01257
L13 QK: 0.01052 0.01184 0.01198 0.00853  OV: 0.01182 0.01274 0.01225 0.01072
L14 QK: 0.00647 0.00717 0.00673 0.00728  OV: 0.00829 0.00915 0.01083 0.00896
L15 QK: 0.01097 0.01269 0.01163 0.01025  OV: 0.01021 0.01130 0.01292 0.01117
L16 QK: 0.01204 0.01258 0.01231 0.01013  OV: 0.01399 0.01555 0.01219 0.01590
L17 QK: 0.00799 0.00971 0.00897 0.00797  OV: 0.00768 0.01132 0.00969 0.00987

Top 16 by sqrt(QK_relative² + OV_relative²):
This is a structural discrepancy score, NOT a behavioural importance score.
layer head QK_rel OV_rel QK_delta OV_delta Q_piece K_piece QK_cos O_piece V_piece OV_cos
    4    1 0.01274 0.01768 0.1599 0.1146 0.11651 0.10446 0.0444 0.098901 0.04417 0.1604
    3    3 0.00947 0.01884 0.181 0.10207 0.13419 0.11405 0.0571 0.091847 0.037042 0.0898
   16    1 0.01258 0.01555 0.16149 0.12031 0.11614 0.10589 0.0560 0.091856 0.070494 0.0824
    8    2 0.01407 0.01415 0.16537 0.11313 0.11282 0.11529 0.0510 0.096367 0.052616 0.0732
    4    0 0.01155 0.01622 0.16203 0.11506 0.1185 0.10677 0.0321 0.1004 0.047612 0.0932
    3    1 0.01355 0.01406 0.16013 0.1068 0.11437 0.10671 0.0481 0.089818 0.048685 0.1109
   11    1 0.01411 0.01277 0.17401 0.13947 0.1113 0.12888 0.0447 0.10712 0.075092 0.1454
   16    3 0.01013 0.01590 0.16939 0.11892 0.11829 0.11214 0.0800 0.095494 0.063502 0.0817
   16    0 0.01204 0.01399 0.17031 0.12173 0.12182 0.11298 0.0510 0.09094 0.076144 0.0542
    4    2 0.01160 0.01425 0.16308 0.11794 0.11854 0.10742 0.0394 0.10079 0.051646 0.1042
   11    0 0.01360 0.01164 0.17758 0.13916 0.11109 0.13468 0.0353 0.096214 0.094274 0.0673
    3    2 0.01231 0.01264 0.1691 0.10891 0.1161 0.11416 0.0786 0.089093 0.05117 0.1433
    5    3 0.01131 0.01328 0.17249 0.13892 0.12069 0.118 0.0444 0.10498 0.082018 0.0900
   13    1 0.01184 0.01274 0.16322 0.13085 0.10891 0.11665 0.0461 0.097935 0.077037 0.1058
   15    2 0.01163 0.01292 0.19222 0.14275 0.12914 0.13426 0.0648 0.099723 0.09311 0.0950
   16    2 0.01231 0.01219 0.1658 0.12545 0.11813 0.10801 0.0734 0.093085 0.075351 0.0994

=== NORMALIZATION PARAMETER RANGES ===
input_layernorm.weight A[min,median,max]= [ -0.8438  12.1875 233.    ] B[min,median,max]= [ -0.8438  12.1875 233.    ] delta_norm= 0.004471951235465239
post_attention_layernorm.weight A[min,median,max]= [-71.5  13.  712. ] B[min,median,max]= [-71.5  13.  712. ] delta_norm= 0.004452117800267246
pre_feedforward_layernorm.weight A[min,median,max]= [-1.0547  3.4297 69.5   ] B[min,median,max]= [-1.0547  3.4297 69.5   ] delta_norm= 0.008490344718207731
post_feedforward_layernorm.weight A[min,median,max]= [-133.     20.25 1656.  ] B[min,median,max]= [-133.     20.25 1656.  ] delta_norm= 0.0058722367288477
self_attn.q_norm.weight A[min,median,max]= [-3.7812  0.7305 16.125 ] B[min,median,max]= [-3.7812  0.7305 16.125 ] delta_norm= 0.006143143279908518
self_attn.k_norm.weight A[min,median,max]= [-3.6875  0.9375 29.5   ] B[min,median,max]= [-3.6875  0.9375 29.5   ] delta_norm= 0.00464327640130778
