OUTPUT of run 2 of at most 60:
```
=== ATTENTION ROUTING COEFFICIENTS ===
Uses Gemma3 RMSNorm learned multiplier (1 + stored weight).
Input RMSNorm's learned diagonal is folded into Q and K.
Query/key RMSNorm learned diagonals are included in RoPE coefficients.
State-dependent RMS denominators are NOT evaluated.
Their weight-derived quadratic forms are compared separately.
No attention logits, positions, hidden states, or probabilities are computed.

Top 24 by relative difference of full rotary coefficient stack:
This is NOT a behavioral-importance ranking.
layer head  route_rel  route_abs  Qgram_rel  Kgram_rel  B/A_size  top_freq  freq_share
   11    0   0.029363     55.039   0.003818   0.007875  0.999099        88     0.25697
   11    1   0.027432     63.305   0.004876   0.007875  0.999269        88     0.17518
   12    0   0.021789     484.54   0.010252   0.005029  0.999893       111     0.07313
   12    2   0.021751     498.04   0.004614   0.005029  0.999877       111     0.05701
   11    3   0.021607     42.022   0.005791   0.007875  0.999973        88     0.14091
   17    1   0.021084     528.13   0.003710   0.002709  0.999146        87     0.17925
   12    3   0.019655     460.05   0.004323   0.005029  1.000179       111     0.04188
    5    3   0.019419      505.6   0.005266   0.002566  1.000464       112     0.05642
   17    0   0.018100     466.07   0.004285   0.002709  1.000044        87     0.19948
   11    2   0.017407     38.091   0.007687   0.007875  0.999656        88     0.14960
   12    1   0.016833     446.32   0.006640   0.005029  1.000162       115     0.04731
   10    1   0.016264     537.24   0.002623   0.004290  1.000371        75     0.09102
    5    0   0.016091     544.67   0.004011   0.002566  1.000040       112     0.05770
    5    2   0.016080     515.51   0.004864   0.002566  1.000306       112     0.06581
   17    2   0.015417     415.88   0.005273   0.002709  0.999638        87     0.12884
   14    1   0.015037     893.23   0.003174   0.003400  1.000099       122     0.06273
   13    3   0.014899     212.07   0.005528   0.002874  1.000033        71     0.07352
   14    0   0.014605     906.93   0.003136   0.003400  0.999999        79     0.04071
    5    1   0.014405     489.83   0.004928   0.002566  1.000159       112     0.08423
   14    3   0.014355      950.5   0.003338   0.003400  0.999908       104     0.04647
   15    2   0.014349      971.8   0.004346   0.004016  1.000295       120     0.07677
   13    1   0.014211     290.07   0.003508   0.002874  1.000144        71     0.06245
   14    2   0.014126     873.29   0.002980   0.003400  1.000087       115     0.06099
   15    0   0.014085     934.14   0.004062   0.004016  0.999951        73     0.07159

Per-layer routing summary:
layer median_route_rel max_route_rel head_at_max median_Qgram_rel Kgram_rel
    0         0.009337      0.009582           3         0.003262  0.003465
    1         0.010711      0.011526           2         0.003628  0.005312
    2         0.009766      0.010514           2         0.004088  0.005379
    3         0.009455      0.013558           1         0.004632  0.004278
    4         0.009999      0.010942           1         0.002956  0.003397
    5         0.016086      0.019419           3         0.004896  0.002566
    6         0.012053      0.013020           2         0.004522  0.003769
    7         0.010548      0.011790           3         0.003743  0.004736
    8         0.009960      0.010997           2         0.004014  0.003928
    9         0.010384      0.011659           0         0.003679  0.004559
   10         0.012391      0.016264           1         0.004282  0.004290
   11         0.024520      0.029363           0         0.005334  0.007875
   12         0.020703      0.021789           0         0.005627  0.005029
   13         0.014104      0.014899           3         0.003590  0.002874
   14         0.014480      0.015037           1         0.003155  0.003400
   15         0.013577      0.014349           2         0.004328  0.004016
   16         0.011672      0.012296           1         0.002464  0.003116
   17         0.016758      0.021084           1         0.004124  0.002709

Relative differences of learned RMSNorm multipliers:
layer input_scale query_scale key_scale
    0    0.000000    0.000019  0.000028
    1    0.000000    0.000000  0.000000
    2    0.000000    0.000038  0.000011
    3    0.000000    0.000053  0.000025
    4    0.000001    0.000042  0.000021
    5    0.000002    0.000020  0.000000
    6    0.000001    0.000094  0.000046
    7    0.000001    0.000041  0.000040
    8    0.000001    0.000040  0.000025
    9    0.000002    0.000068  0.000035
   10    0.000000    0.000058  0.000019
   11    0.000007    0.000032  0.000010
   12    0.000001    0.000032  0.000022
   13    0.000004    0.000066  0.000030
   14    0.000003    0.000034  0.000037
   15    0.000001    0.000034  0.000046
   16    0.000000    0.000038  0.000031
   17    0.000000    0.000037  0.000019

Maximum relative error in explicit coefficient verification: 2.7184969195200045e-11

B has larger coefficient-stack norm in 40 of 72 heads.
Larger routing coefficients do not imply more user agreement:
normalization, softmax, attended content, and downstream decoding matter.

```