=== LEARNED NORMALIZATION SCALES ===
RMS across parameter coordinates of (1 + weight), not hidden-state RMS.
layer  input_A  postAttn_A  preMLP_A  postMLP_A  qNorm_A  kNorm_A  max_relative_AB
    0      14.55      2.396      6.132      6.865      2.119      2.146     2.57146e-05
    1      26.73      9.365      31.32      9.372      2.994      2.898     1.18812e-05
    2      32.79      5.446      25.16      14.16      1.871      2.747     3.36226e-05
    3      30.06      12.53      32.24      35.57      1.658      1.963     6.93025e-05
    4       50.3      13.17      17.01      28.05      1.915      2.933     5.78075e-05
    5      17.93      28.76      15.18      34.17       2.32      3.987     3.05356e-05
    6       27.4      19.14      9.845      35.87      1.579      2.204     6.84586e-05
    7      25.96      21.14      9.025      33.88          2       2.49     4.65754e-05
    8       26.1      30.62      8.426      37.79      2.021      2.437     4.52534e-05
    9      22.48      32.78      5.509       43.1      2.054       2.47     5.47819e-05
   10       25.5      57.61      7.305      74.78       1.82      3.702     7.23318e-05
   11      9.733      185.2      9.469      147.5      1.859      4.787      3.7669e-05
   12      31.13      201.8      6.029      207.1      2.003      2.767     3.62398e-05
   13      21.36      166.2      5.767      215.1      1.851      2.381     4.72146e-05
   14      30.13      188.8      5.388      246.6      2.249      3.204     4.07494e-05
   15      38.58      246.2      3.995      292.5      2.569       2.53     4.80569e-05
   16      39.69      163.4       3.89      340.3      2.262      2.534     3.71035e-05
   17      23.13      157.2      4.204      376.5      2.538      3.059     3.11666e-05

=== HEAD-LEVEL COMPOSED DIFFERENCES ===
All differences use B-A. 'id' is Frobenius cosine with identity.
rw_cos compares the two OV change contributions; it is not an activation statistic.
 L H  rawOV_rel scaledOV_rel   QK_rel     OV_id_A    OV_id_B   delta_id    rw_cos
 0 0  0.0076452     0.009225   0.012181    0.019092    0.019095  -0.000480   0.028787
 0 1   0.011057     0.021705  0.0091683   -0.044610   -0.044712  -0.000413  -0.018415
 0 2   0.010037    0.0074203  0.0079705   -0.004200   -0.004196   0.000348  -0.019266
 0 3  0.0073909    0.0076407   0.010395    0.009956    0.009946   0.000081   0.144875
 1 0   0.012916     0.010858  0.0094948    0.007646    0.007649   0.000143   0.177928
 1 1    0.01209     0.010648  0.0097521    0.002143    0.002138  -0.000535   0.144691
 1 2   0.011898    0.0085975   0.012666    0.010430    0.010427  -0.000182   0.165357
 1 3   0.010541    0.0083477  0.0097506   -0.000684   -0.000684  -0.000063   0.060517
 2 0   0.010138     0.012261   0.010914    0.011544    0.011536  -0.000016   0.135485
 2 1  0.0099463     0.012229  0.0087578   -0.017657   -0.017653  -0.000052   0.077881
 2 2  0.0089989     0.010282   0.012072   -0.006995   -0.006999  -0.000389   0.067725
 2 3   0.012961     0.020124   0.014433   -0.002755   -0.002757  -0.000144   0.045755
 3 0   0.011529     0.015128   0.008673   -0.025371   -0.025379  -0.000264   0.140648
 3 1    0.01406     0.019943   0.018886    0.013834    0.013854  -0.000244   0.160109
 3 2   0.012643      0.01508   0.012594    0.048501    0.048496  -0.000489   0.105403
 3 3   0.018836     0.022078  0.0065247   -0.008468   -0.008477  -0.000150   0.086862
 4 0   0.016224     0.021169    0.01227   -0.000130   -0.000123   0.000343   0.122975
 4 1   0.017684     0.031444   0.017382    0.006639    0.006661   0.000571   0.166536
 4 2   0.014252     0.021399   0.014718   -0.001902   -0.001886   0.000791   0.237014
 4 3   0.010726     0.013634   0.013178   -0.007910   -0.007893   0.000765   0.154519
 5 0  0.0087974    0.0075906   0.022973    0.008309    0.008287  -0.001830   0.142374
 5 1   0.011443     0.015018   0.014135    0.002407    0.002413   0.000359   0.134720
 5 2    0.01261     0.014438   0.018361    0.008514    0.008518  -0.000072   0.108379
 5 3   0.013281      0.01183   0.024607   -0.011512   -0.011515  -0.000651   0.013571
 6 0   0.011131     0.017484   0.015127    0.063028    0.063097   0.000911   0.086220
 6 1  0.0092194     0.011857   0.017624   -0.063850   -0.063798  -0.000096   0.005777
 6 2   0.012107     0.017112   0.023186    0.095339    0.095374  -0.000376   0.178876
 6 3   0.010894      0.01398   0.016295    0.059677    0.059609  -0.000351   0.106377
 7 0   0.011922     0.014711   0.012524    0.121132    0.121182  -0.000410   0.165980
 7 1  0.0087306    0.0098073   0.016732   -0.149314   -0.149476  -0.000657   0.043078
 7 2   0.010228     0.011131   0.014572    0.079967    0.080012  -0.000078   0.088307
 7 3    0.01365     0.018261   0.020518    0.072678    0.072631  -0.000548   0.135438
 8 0   0.011199     0.013034    0.01271   -0.111416   -0.111395   0.000247   0.189663
 8 1   0.012603     0.017408     0.0117    0.091480    0.091442   0.000497   0.068082
 8 2   0.014148     0.019953   0.016382    0.134646    0.134668   0.000616   0.016746
 8 3     0.0108     0.015121   0.013557    0.032520    0.032528   0.000698   0.071227
 9 0   0.011732     0.013813   0.018547    0.039620    0.039613  -0.000122   0.167082
 9 1  0.0086569     0.011088   0.013897   -0.142374   -0.142390   0.000089   0.019524
 9 2   0.011157     0.014154   0.018675    0.152314    0.152336   0.000733   0.110976
 9 3   0.011223     0.013467   0.012444    0.068236    0.068282  -0.000811   0.062306
10 0   0.010734     0.010902   0.022533    0.070908    0.070900  -0.001261   0.204984
10 1   0.010347     0.011986   0.032764    0.011383    0.011381  -0.000109   0.144256
10 2   0.012432     0.014925   0.010355    0.054897    0.054904   0.000167   0.153407
10 3  0.0082558     0.007634   0.017812   -0.016629   -0.016627   0.000124   0.080658
11 0    0.01164     0.012294   0.046586    0.002534    0.002533  -0.000113   0.098112
11 1   0.012774     0.020808   0.038064    0.053633    0.053632   0.000087   0.136672
11 2  0.0091472     0.010203   0.018113    0.006717    0.006711  -0.000708   0.206157
11 3  0.0082757    0.0085605   0.027372   -0.011000   -0.010995   0.000511  -0.086456
12 0   0.011856     0.014131   0.030242   -0.006842   -0.006840   0.000211   0.196665
12 1   0.011047     0.012895   0.023867    0.001764    0.001765   0.000056   0.302121
12 2  0.0090539    0.0089586   0.031127    0.009830    0.009829   0.000037   0.269402
12 3   0.012573      0.01272   0.025832    0.008084    0.008086  -0.000041   0.205238
13 0   0.011817     0.010513   0.016879    0.010266    0.010266  -0.000127   0.098269
13 1   0.012744     0.012921   0.021734    0.015603    0.015607   0.000143   0.087524
13 2   0.012248      0.01355   0.023653    0.039707    0.039730   0.000362   0.030823
13 3   0.010716    0.0083401      0.012   -0.001036   -0.001036  -0.000058   0.254258
14 0  0.0082921    0.0082098   0.017395   -0.005483   -0.005478   0.000571   0.036690
14 1  0.0091456    0.0093077   0.025863    0.061875    0.061866  -0.000237   0.062844
14 2   0.010826     0.010506   0.017918    0.019606    0.019604  -0.000153   0.098954
14 3  0.0089603    0.0088237   0.020009    0.011015    0.011011  -0.000192   0.072966
15 0   0.010206     0.012147   0.022712    0.059036    0.059030   0.000027   0.097709
15 1     0.0113     0.016637   0.014591   -0.096406   -0.096395   0.000211   0.029402
15 2   0.012915     0.018849   0.018811    0.058147    0.058144   0.000122   0.172451
15 3   0.011166     0.015267    0.01336    0.067654    0.067633   0.000256   0.159178
16 0   0.013992     0.012581   0.014642    0.008113    0.008115   0.000181   0.057959
16 1   0.015549     0.014362   0.016645    0.025494    0.025482  -0.000237   0.079311
16 2    0.01219     0.012147   0.016013    0.045774    0.045773  -0.000093   0.127097
16 3   0.015897     0.014269    0.01167    0.005337    0.005338   0.000107   0.082181
17 0  0.0076799       0.0069   0.015541    0.013785    0.013789   0.000483   0.031972
17 1   0.011319     0.014634   0.022978    0.107120    0.107108  -0.000025   0.217987
17 2  0.0096863    0.0090783  0.0089019    0.000935    0.000943   0.000788   0.089947
17 3  0.0098722     0.010881   0.014888   -0.006029   -0.006031  -0.000154   0.032470

=== LARGEST SCALED OV RELATIVE DIFFERENCES ===
L04H1 relative=0.0314439 absolute=86.9788 reinforcement=1.09557 identity_alignment_change=+0.0000218
L03H3 relative=0.0220783 absolute=41.5135 reinforcement=1.07019 identity_alignment_change=-0.0000084
L00H1 relative=0.0217048 absolute=3.66843 reinforcement=0.990618 identity_alignment_change=-0.0001021
L04H2 relative=0.0213994 absolute=94.4952 reinforcement=1.14809 identity_alignment_change=+0.0000159
L04H0 relative=0.0211688 absolute=79.7112 reinforcement=1.08659 identity_alignment_change=+0.0000068
L11H1 relative=0.020808 absolute=233.797 reinforcement=1.07314 identity_alignment_change=-0.0000004
L02H3 relative=0.0201238 absolute=21.4519 reinforcement=1.0326 identity_alignment_change=-0.0000023
L08H2 relative=0.0199525 absolute=80.9438 reinforcement=1.01092 identity_alignment_change=+0.0000224
L03H1 relative=0.0199428 absolute=44.6888 reinforcement=1.126 identity_alignment_change=+0.0000205
L15H2 relative=0.0188486 absolute=1369.14 reinforcement=1.17236 identity_alignment_change=-0.0000023
L07H3 relative=0.0182613 absolute=74.7313 reinforcement=1.1354 identity_alignment_change=-0.0000472
L06H0 relative=0.0174839 absolute=61.8483 reinforcement=1.07594 identity_alignment_change=+0.0000695

=== LARGEST QK NUMERATOR RELATIVE DIFFERENCES ===
L11H0 relative=0.0465857 QK_identity_A=+0.018163 QK_identity_B=+0.018535
L11H1 relative=0.0380637 QK_identity_A=+0.036234 QK_identity_B=+0.036300
L10H1 relative=0.0327639 QK_identity_A=-0.026533 QK_identity_B=-0.026155
L12H2 relative=0.0311272 QK_identity_A=+0.096740 QK_identity_B=+0.096951
L12H0 relative=0.0302416 QK_identity_A=+0.053999 QK_identity_B=+0.054115
L11H3 relative=0.0273716 QK_identity_A=-0.019503 QK_identity_B=-0.019302
L14H1 relative=0.025863 QK_identity_A=+0.043085 QK_identity_B=+0.043302
L12H3 relative=0.0258321 QK_identity_A=+0.036582 QK_identity_B=+0.036471
L05H3 relative=0.0246073 QK_identity_A=+0.173952 QK_identity_B=+0.174187
L12H1 relative=0.0238667 QK_identity_A=+0.083399 QK_identity_B=+0.083392
L13H2 relative=0.0236531 QK_identity_A=+0.037655 QK_identity_B=+0.037328
L06H2 relative=0.0231859 QK_identity_A=-0.046529 QK_identity_B=-0.046628

=== EXACT SINGULAR VALUES OF SIX LEADING OV DIFFERENCES ===
Selection is by relative operator difference, not behavioral importance.
L04H1 s1=69.8708 top1_energy=0.64531 top8_energy=0.95054 participation=2.12 k50=1 k90=3
L03H3 s1=33.3578 top1_energy=0.64568 top8_energy=0.84719 participation=2.35 k50=1 k90=14
L00H1 s1=3.31939 top1_energy=0.81876 top8_energy=0.92912 participation=1.49 k50=1 k90=5
L04H2 s1=80.5522 top1_energy=0.72667 top8_energy=0.96329 participation=1.80 k50=1 k90=3
L04H0 s1=62.8552 top1_energy=0.62179 top8_energy=0.95405 participation=2.16 k50=1 k90=3
L11H1 s1=72.4849 top1_energy=0.09612 top8_energy=0.55462 participation=20.18 k50=7 k90=27

=== DIRECTIONAL SUMMARY ===
OV identity alignment B-A positive_heads= 36 negative_heads= 36 median= 1.9993305578583538e-07 range= (-0.00016239883925914245, 6.951886989163547e-05)
QK identity alignment B-A positive_heads= 41 negative_heads= 31 median= 2.2500490990460045e-05 range= (-0.00032749441710559074, 0.0003780644330299253)

Limits: OV omits attention routing and state-dependent normalization.
QK omits state-dependent denominators, nonzero relative-position RoPE,
masking, and softmax. Neither operator identifies a user's guess, truth,
or an answer token. No model or submodule was run on an input.
