=== NORM DIAGONAL GAINS ===
Uses gain=1+weight. RMS below is a statistic of weights, not a forward pass.
layer | RMS gains in order: input, post-attn, pre-MLP, post-MLP, q, k | differing norm entries
 0 |      14.55      2.396      6.132      6.865      2.119      2.146 | [0, 2, 0, 104, 5, 6]
 1 |      26.73      9.365      31.32      9.372      2.994      2.898 | [0, 25, 0, 43, 0, 0]
 2 |      32.79      5.446      25.16      14.16      1.871      2.747 | [1, 46, 1, 6, 4, 6]
 3 |      30.06      12.53      32.24      35.57      1.658      1.963 | [2, 15, 4, 10, 14, 9]
 4 |       50.3      13.17      17.01      28.05      1.915      2.933 | [4, 7, 10, 2, 18, 10]
 5 |      17.93      28.76      15.18      34.17       2.32      3.987 | [6, 2, 11, 0, 3, 0]
 6 |       27.4      19.14      9.845      35.87      1.579      2.204 | [4, 6, 11, 0, 26, 12]
 7 |      25.96      21.14      9.025      33.88          2       2.49 | [1, 0, 18, 0, 14, 12]
 8 |       26.1      30.62      8.426      37.79      2.021      2.437 | [3, 4, 16, 0, 10, 11]
 9 |      22.48      32.78      5.509       43.1      2.054       2.47 | [6, 5, 15, 0, 20, 10]
10 |       25.5      57.61      7.305      74.78       1.82      3.702 | [0, 3, 19, 0, 16, 9]
11 |      9.733      185.2      9.469      147.5      1.859      4.787 | [10, 0, 8, 0, 7, 5]
12 |      31.13      201.8      6.029      207.1      2.003      2.767 | [5, 0, 13, 0, 7, 2]
13 |      21.36      166.2      5.767      215.1      1.851      2.381 | [34, 0, 37, 0, 8, 6]
14 |      30.13      188.8      5.388      246.6      2.249      3.204 | [19, 0, 74, 0, 7, 9]
15 |      38.58      246.2      3.995      292.5      2.569       2.53 | [1, 0, 29, 0, 11, 14]
16 |      39.69      163.4       3.89      340.3      2.262      2.534 | [0, 0, 24, 0, 12, 7]
17 |      23.13      157.2      4.204      376.5      2.538      3.059 | [0, 1, 27, 0, 9, 10]

=== PER-LAYER COMPOSED DIFFERENCES ===
These are sums of squared head-map differences, not behavioral effects.
layer | raw_OV_delta_RSS gain_OV_delta_RSS raw_QK_delta_RSS gain_QK_delta_RSS
 0 |          0.272939           7.54983          0.288079            237.51
 1 |           0.23661           90.6535          0.280219           1238.34
 2 |          0.254031           49.1667          0.304051           985.985
 3 |          0.212776           91.5756          0.313861           688.375
 4 |          0.230294           188.879           0.31948           1417.93
 5 |          0.273668           184.108          0.302952           726.105
 6 |          0.262492           127.171          0.350203           513.028
 7 |          0.244687            146.17          0.339808           590.266
 8 |          0.234115           198.862          0.326051           477.965
 9 |          0.291318           194.744          0.334419           491.103
10 |          0.321785           459.656          0.367183             644.4
11 |          0.289714           676.283          0.332431           75.0225
12 |          0.315095           2120.57          0.317789           659.174
13 |          0.262011           1086.03          0.326377           387.429
14 |          0.279565           1569.88          0.335646           1317.88
15 |           0.29468           2774.79          0.361218           1055.92
16 |          0.251945           1510.08          0.329904           1851.24
17 |          0.241875           988.954          0.334072            673.94

=== HEADS: LARGEST GAIN-WEIGHTED OV DIFFERENCES ===
layer head | ||map_A|| ||map_B|| ||delta|| relative_delta | raw_relative_delta | identity_cos_A identity_cos_B
15  1 |       99093       99101      1636.5  0.0165139 |   0.011054 | -0.0964304 -0.0964197
15  2 |       72642       72621      1373.9  0.0189165 |  0.0130853 | +0.0581424 +0.0581651
15  3 |       86544       86553      1354.6  0.0156513 |  0.0112672 | +0.0676462 +0.0676387
12  1 |       97757       97751        1241  0.0126949 |  0.0111012 | +0.0017616 +0.0017635
12  2 |  1.2049e+05  1.2051e+05      1159.4 0.00962141 | 0.00945398 | +0.0098289 +0.0098291
15  0 |       86722       86727      1139.6  0.0131408 |  0.0106151 | +0.0590363 +0.0590339
12  3 |       73572       73572       945.3  0.0128486 |  0.0125873 | +0.0080872 +0.0080852
14  3 |       92214       92217      876.01 0.00949955 | 0.00959746 | +0.0110177 +0.0110177
12  0 |       61624       61630      847.98    0.01376 |  0.0114303 | -0.0068407 -0.0068386
14  2 |       80635       80634      793.92 0.00984587 |   0.010205 | +0.0196018 +0.0196012
16  2 |       61604       61607       773.2  0.0125508 |  0.0123612 | +0.0457768 +0.0457678
16  0 |       56618       56624      766.22  0.0135324 |  0.0149753 | +0.0081153 +0.0081132
16  1 |       48255       48256      753.96  0.0156244 |  0.0164674 | +0.0255015 +0.0254947
14  1 |       79294       79297      742.59 0.00936487 | 0.00915899 | +0.0618775 +0.0618751
16  3 |       50028       50027      725.91  0.0145102 |  0.0158639 | +0.0053431 +0.0053378
14  0 |       85650       85643      717.89 0.00838204 | 0.00807401 | -0.0054809 -0.0054838

=== HEADS: LARGEST GAIN-WEIGHTED QK NUMERATOR DIFFERENCES ===
layer head | ||map_A|| ||map_B|| ||delta|| relative_delta | raw_relative_delta | identity_cos_A identity_cos_B
16  0 |       58714       58700      973.23  0.0165777 |  0.0123259 | +0.0706427 +0.0706372
16  1 |       52641       52651      963.12  0.0182943 |  0.0131276 | -0.0165160 -0.0164274
16  3 |       74131       74154      887.26   0.011967 | 0.00963537 | +0.1191197 +0.1191616
16  2 |       55904       55899      874.69  0.0156469 |   0.011528 | +0.0967017 +0.0969206
 4  0 |       57782       57793      739.32  0.0127939 |  0.0117594 | +0.0209754 +0.0209925
14  3 |       30636       30653      727.23  0.0237312 | 0.00759912 | +0.0791008 +0.0792944
 4  2 |       48514       48525      717.72  0.0147924 |  0.0114002 | -0.0229495 -0.0228611
 1  3 |       60716       60709      696.45  0.0114713 | 0.00918227 | -0.0186618 -0.0187069
14  0 |       29882       29924      693.52  0.0231924 | 0.00698608 | +0.0280684 +0.0285252
 4  1 |       38437       38418      690.13  0.0179592 |  0.0128116 | -0.0751997 -0.0751815
 4  3 |       55109       55107       687.4  0.0124738 |  0.0116657 | +0.0133235 +0.0131876
14  1 |       24210       24189      632.96   0.026156 | 0.00715394 | +0.0433107 +0.0433480
 1  0 |       54575       54572      617.71  0.0113189 | 0.00826823 | -0.0392036 -0.0391525
 1  2 |       42757       42753      588.53  0.0137651 | 0.00707491 | +0.0594888 +0.0592965
14  2 |       28748       28770      571.26  0.0198639 | 0.00703056 | +0.0556465 +0.0557555
 1  1 |       65257       65254      566.14  0.0086757 | 0.00755166 | +0.0112169 +0.0111427

=== LARGEST CHANGES IN OV IDENTITY ALIGNMENT ===
Positive difference means B is more identity-aligned; not evidence of user-guess copying.
layer head | gain_identity_cos_A gain_identity_cos_B change | raw_identity_cos_A raw_identity_cos_B
 7  3 | +0.07258347 +0.07276574 +0.00018227 | +0.20942691 +0.20943237
 7  0 | +0.12095417 +0.12113285 +0.00017868 | +0.33474171 +0.33477045
 8  0 | -0.11145430 -0.11133612 +0.00011818 | -0.24326021 -0.24323366
 9  2 | +0.15230994 +0.15242344 +0.00011350 | +0.24116306 +0.24117465
 3  0 | -0.02539289 -0.02534276 +0.00005014 | -0.15324948 -0.15325716
 3  2 | +0.04853142 +0.04848296 -0.00004846 | +0.26541113 +0.26540492
 9  3 | +0.06829110 +0.06825141 -0.00003969 | +0.11637567 +0.11640176
 7  1 | -0.14937099 -0.14940934 -0.00003835 | -0.44306249 -0.44306799
 6  1 | -0.06379356 -0.06375818 +0.00003537 | -0.13987176 -0.13985805
 6  2 | +0.09539629 +0.09536621 -0.00003008 | +0.19364982 +0.19364660
 3  1 | +0.01387397 +0.01384564 -0.00002833 | +0.06613625 +0.06612688
 9  0 | +0.03961549 +0.03964277 +0.00002728 | +0.06346766 +0.06347702
11  1 | +0.05365777 +0.05363249 -0.00002528 | +0.40658577 +0.40658781
 8  2 | +0.13459636 +0.13462148 +0.00002512 | +0.16910003 +0.16908208
 5  1 | +0.00241905 +0.00239538 -0.00002366 | +0.03951776 +0.03949001
 4  1 | +0.00664997 +0.00662653 -0.00002343 | +0.07663681 +0.07661722

Interpretation limits:
- OV maps omit input-dependent attention probabilities and RMS denominators.
- QK maps omit RMS denominators and nonzero-offset rotary rotations.
- Shared K and V mean head differences are not independent parameter changes.
- No tokens, test vectors, hidden states, activations, or forward passes were used.
