OUTPUT of run 9 of at most 60:
```
=== NEW HYPOTHESIS: CROSS-HEAD CANCELLATION / REDISTRIBUTION ===
M_h = O_h @ V, with learned diagonals additionally included in scaled results.
Common coefficient: C = sum_h M_h / sqrt(4).
Three orthogonal contrast coefficients complete an exact decomposition.
Sum of squared coefficient norms is preserved by this head-index transform.
These are coefficient identities, not evaluations with equal attention patterns.
Changes in attention patterns themselves are outside this decomposition.
For independently randomized whole-head signs, expected common energy fraction is 1/4.
That is a geometric reference, not a null model of fine-tuning.

=== COMMON VERSUS HEAD-CONTRAST DIFFERENCES ===
Percentages refer to squared coefficient norms, not behavioural contributions.
layer raw_common_delta% scaled_common_delta% scaled_baseline_common% scaled_common_rel scaled_contrast_rel
    0    41.539    47.376    35.277 0.010598 0.008246
    1    46.677    55.325    40.422 0.010937 0.008095
    2    42.085    49.120    17.380 0.021091 0.009845
    3    49.041    53.288    22.283 0.026862 0.013467
    4    60.805    72.173    21.274 0.034665 0.011189
    5    47.836    58.790    41.574 0.012333 0.008710
    6    47.370    52.521     9.527 0.033704 0.010399
    7    41.110    47.930    10.208 0.026049 0.009155
    8    52.398    48.581    16.896 0.024871 0.011537
    9    41.657    45.687     8.058 0.030301 0.009781
   10    48.107    52.344    26.360 0.014241 0.008130
   11    35.385    29.372     2.958 0.031958 0.008652
   12    49.334    57.079    42.012 0.013384 0.009878
   13    44.766    42.507    18.544 0.014950 0.008296
   14    40.282    36.764    16.819 0.013590 0.008014
   15    52.055    36.629     8.878 0.031948 0.013117
   16    66.119    59.910    18.075 0.024007 0.009224
   17    38.474    38.743     9.584 0.018585 0.007608

=== COORDINATION OF THE FOUR HEAD DIFFERENCES, SCALED ===
The leading mode is learned from the 4x4 Gram matrix of weight differences.
layer top_headmode_energy% topmode_common_alignment% delta_pair_cos[min,median,max] leading_headmode[H0,H1,H2,H3]
    0    48.402    96.324 [0.1297,0.2837,0.4974] [0.6251, 0.537, 0.3898, 0.4109]
    1    56.394    97.585 [0.3319,0.4059,0.5312] [0.3865, 0.5874, 0.5444, 0.4574]
    2    50.037    96.892 [0.1994,0.2973,0.4488] [0.3621, 0.6023, 0.5324, 0.4719]
    3    54.977    95.966 [0.2892,0.3776,0.5006] [0.5919, 0.5844, 0.3631, 0.4198]
    4    72.922    98.839 [0.5097,0.6548,0.6990] [0.4355, 0.4538, 0.5647, 0.5343]
    5    58.898    99.768 [0.2981,0.4619,0.6470] [0.516, 0.4579, 0.51, 0.5138]
    6    53.384    96.836 [0.1469,0.3816,0.6134] [0.5701, 0.3422, 0.544, 0.5117]
    7    48.094    99.527 [0.1215,0.3407,0.5790] [0.5264, 0.4755, 0.4554, 0.5379]
    8    54.990    70.426 [0.0220,0.4453,0.6966] [0.8784, 0.3559, 0.2617, 0.1825]
    9    45.887    99.000 [0.1005,0.2540,0.5907] [0.4738, 0.4271, 0.5497, 0.5394]
   10    53.064    98.296 [0.2732,0.3346,0.4921] [0.5792, 0.4026, 0.4746, 0.5265]
   11    54.502     4.185 [-0.4203,0.2373,0.4151] [0.0842, 0.0861, 0.8112, -0.5723]
   12    61.042    92.002 [0.1728,0.4842,0.5851] [0.3266, 0.7106, 0.4494, 0.4317]
   13    50.241    69.151 [0.0916,0.2589,0.4291] [0.3549, 0.1827, 0.2409, 0.8847]
   14    37.726    94.099 [0.1025,0.1544,0.2119] [0.361, 0.4034, 0.6768, 0.4989]
   15    52.502    26.282 [-0.1959,0.1601,0.7248] [0.3468, -0.4697, 0.5748, 0.5734]
   16    60.132    99.519 [0.3609,0.4634,0.5648] [0.458, 0.513, 0.5481, 0.4761]
   17    40.559    91.206 [0.0656,0.1491,0.3501] [0.4749, 0.6669, 0.5157, 0.2526]

=== COMMON-COMPONENT ATTRIBUTION AND QUERY-MATRIX DISTINCTNESS ===
Query cosines are matrix cosines only; no query vectors are evaluated.
layer common_output_side_share% common_value_side_share% query_matrix_pair_cos[min,max]
    0    74.146    25.854 [-0.07328,0.25152]
    1    35.849    64.151 [0.06934,0.19889]
    2    74.942    25.058 [-0.01231,0.11411]
    3    84.230    15.770 [-0.36140,0.16506]
    4    87.290    12.710 [0.03620,0.14286]
    5    44.081    55.919 [0.18614,0.41901]
    6    90.077     9.923 [0.20313,0.46788]
    7    60.722    39.278 [0.12174,0.49362]
    8    80.912    19.088 [0.14114,0.38764]
    9    89.528    10.472 [0.23005,0.57086]
   10    77.263    22.737 [0.03605,0.49229]
   11    95.658     4.342 [0.04853,0.39759]
   12    58.182    41.818 [0.08663,0.45546]
   13    74.229    25.771 [0.02901,0.19658]
   14    73.415    26.585 [0.09886,0.22835]
   15    88.770    11.230 [0.15796,0.37996]
   16    81.933    18.067 [0.08298,0.23516]
   17    86.883    13.117 [0.16972,0.54947]

Layers with <10% scaled difference energy in common component: []
Layers with >50% scaled difference energy in common component: [1, 3, 4, 5, 6, 10, 12, 16]
These thresholds are descriptive, not significance tests.
Maximum relative Parseval error: 4.36045056538348e-15
Maximum relative midpoint reconstruction error: 1.0760696879899096e-15
Actual output cancellation cannot be inferred without the attention patterns.

```