OUTPUT of run 14 of at most 60:
```
=== COMMON TRANSPORT VERSUS HEAD-CONTRAST DIFFERENCES ===
Common mode is sum_h C_h / sqrt(H); contrasts sum to zero over heads.
Only coefficient tensors are decomposed; no attention routing is evaluated.
Independent head-sign controls preserve each difference operator's spectrum.
Their exact mean common-energy fraction is 1/H = 0.25
 L rawCommon% scaledCommon% sign_control%[min,median,max] relativeCommon relativeContrast
 0    41.5393       47.3761 [ 11.8719, 24.2770, 47.3761]       0.010598         0.008246
 1    46.6769       55.3246 [ 14.1147, 22.3391, 55.3246]       0.010937         0.008095
 2    42.0851       49.1197 [ 15.6970, 22.4339, 49.1197]       0.021091         0.009845
 3    49.0411       53.2878 [ 15.2502, 21.4438, 53.2878]       0.026862         0.013467
 4    60.8051       72.1734 [  7.2358, 21.5310, 72.1734]       0.034665         0.011189
 5    47.8357       58.7905 [ 10.9041, 21.8534, 58.7905]       0.012333         0.008710
 6    47.3702       52.5208 [ 14.2608, 20.4527, 52.5208]       0.033704         0.010399
 7    41.1098       47.9303 [ 16.3904, 21.0590, 47.9303]       0.026049         0.009155
 8    52.3985       48.5812 [ 13.6988, 20.9985, 48.5812]       0.024871         0.011537
 9    41.6573       45.6875 [ 14.9041, 21.3887, 45.6875]       0.030301         0.009781
10    48.1067       52.3442 [ 12.2511, 22.6608, 52.3442]       0.014241         0.008130
11    35.3854       29.3721 [ 10.0139, 29.5824, 38.3022]       0.031958         0.008652
12    49.3345       57.0794 [ 11.3763, 21.5513, 57.0794]       0.013384         0.009878
13    44.7655       42.5067 [ 14.9872, 24.4657, 42.5067]       0.014950         0.008296
14    40.2817       36.7640 [ 19.7109, 24.3065, 36.7640]       0.013590         0.008014
15    52.0553       36.6291 [ 13.6333, 19.4892, 51.0148]       0.031948         0.013117
16    66.1195       59.9104 [ 13.2211, 22.3419, 59.9104]       0.024007         0.009224
17    38.4744       38.7426 [ 15.5833, 24.8554, 38.7426]       0.018585         0.007608

=== BASELINE COMMON-MODE SHARES AND SIGNED CHANGES ===
These are coefficient-space quantities, not attention/output statistics.
 L commonShare_A commonShare_B commonNorm_B/A differenceCommon_percentile
 0      0.352781      0.352757      1.0002935                      1.0000
 1      0.404511      0.403922      0.9993837                      1.0000
 2      0.173765      0.173830      1.0004468                      1.0000
 3      0.223065      0.222588      0.9983906                      1.0000
 4      0.212585      0.212886      1.0004145                      1.0000
 5      0.415453      0.416020      1.0012757                      1.0000
 6      0.095160      0.095384      1.0015762                      1.0000
 7      0.102435      0.101723      0.9958240                      1.0000
 8      0.169052      0.168878      0.9996640                      1.0000
 9      0.080639      0.080526      0.9990968                      1.0000
10      0.263572      0.263618      1.0001192                      1.0000
11      0.029580      0.029576      0.9999132                      0.5000
12      0.420112      0.420122      0.9999215                      1.0000
13      0.185504      0.185369      0.9996234                      1.0000
14      0.168195      0.168183      1.0000696                      1.0000
15      0.088733      0.088822      1.0006595                      0.8750
16      0.180755      0.180741      1.0000769                      1.0000
17      0.095827      0.095844      1.0000794                      1.0000

=== LEADING HEAD-SPACE MODES OF THE DIFFERENCE ===
Mode signs have an arbitrary overall orientation; relative signs indicate
reinforcing versus opposing operator differences, not semantic polarity.
 L leading_energy_share common_mode_overlap head_coefficients
 0             0.484022            0.963239 [0.6251 0.537  0.3898 0.4109]
 1             0.563944            0.975854 [0.3865 0.5874 0.5444 0.4574]
 2             0.500367            0.968915 [0.3621 0.6023 0.5324 0.4719]
 3             0.549769            0.959659 [0.5919 0.5844 0.3631 0.4198]
 4             0.729225            0.988388 [0.4355 0.4538 0.5647 0.5343]
 5             0.588982            0.997681 [0.516  0.4579 0.51   0.5138]
 6             0.533839            0.968363 [0.5701 0.3422 0.544  0.5117]
 7             0.480937            0.995275 [0.5264 0.4755 0.4554 0.5379]
 8             0.549896            0.704256 [0.8784 0.3559 0.2617 0.1825]
 9             0.458874            0.989999 [0.4738 0.4271 0.5497 0.5394]
10             0.530645            0.982962 [0.5792 0.4026 0.4746 0.5265]
11             0.545025            0.041851 [ 0.0842  0.0861  0.8112 -0.5723]
12             0.610422            0.920016 [0.3266 0.7106 0.4494 0.4317]
13             0.502407            0.691507 [0.3549 0.1827 0.2409 0.8847]
14             0.377255            0.940994 [0.361  0.4034 0.6768 0.4989]
15             0.525016            0.262818 [ 0.3468 -0.4697  0.5748  0.5734]
16             0.601324            0.995186 [0.458  0.513  0.5481 0.4761]
17             0.405589            0.912057 [0.4749 0.6669 0.5157 0.2526]

=== OUTPUT-SIDE / VALUE-SIDE / CROSS-TERM DECOMPOSITION ===
Scaled output-side includes post-attention diagonal-scale changes.
Scaled value-side includes input diagonal-scale changes.
 L common_parts%[O,V,cross] contrast_parts%[O,V,cross]
 0 [73.1081 24.8163  2.0756] [61.4017 34.8936  3.7048]
 1 [29.3759 57.6789 12.9452] [48.7528 40.7407 10.5065]
 2 [69.486  19.6015 10.9124] [57.569  39.3116  3.1193]
 3 [78.7664 10.3058 10.9277] [57.4017 34.3637  8.2346]
 4 [80.7955  6.216  12.9885] [69.2235 27.331   3.4455]
 5 [37.978  49.8151 12.2069] [29.0408 68.0983  2.8609]
 6 [86.8845  6.7309  6.3846] [35.7138 56.4303  7.8559]
 7 [53.3552 31.912  14.7328] [32.1928 64.3693  3.4378]
 8 [79.1155 17.2917  3.5929] [23.9162 60.8262 15.2576]
 9 [86.5101  7.4533  6.0367] [28.2607 63.7719  7.9674]
10 [69.9011 15.3756 14.7233] [41.477  49.9065  8.6165]
11 [93.9143  2.5982  3.4875] [26.5677 66.9483  6.4839]
12 [44.9931 28.6289 26.378 ] [56.0454 33.1504 10.8043]
13 [67.4952 19.0365 13.4683] [44.3524 44.0185 11.6291]
14 [69.8381 23.0079  7.1539] [36.7104 57.042   6.2476]
15 [84.6434  7.1025  8.254 ] [19.7928 70.5625  9.6447]
16 [77.9216 14.0552  8.0232] [41.4758 51.8241  6.7001]
17 [83.218   9.4514  7.3306] [48.4098 43.6695  7.9207]

=== SUMMARY ===
Scaled common fraction min/median/max: [0.2937 0.4885 0.7217]
Layers below the sign-control mean of 25%: []
Layers above the sign-control mean of 25%: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]
Largest contrast-relative differences:
L03: contrast_relative=0.013467, common_relative=0.026862, contrast_energy_share=0.467122
L15: contrast_relative=0.013117, common_relative=0.031948, contrast_energy_share=0.633709
L08: contrast_relative=0.011537, common_relative=0.024871, contrast_energy_share=0.514188
L04: contrast_relative=0.011189, common_relative=0.034665, contrast_energy_share=0.278266
L06: contrast_relative=0.010399, common_relative=0.033704, contrast_energy_share=0.474792
L12: contrast_relative=0.009878, common_relative=0.013384, contrast_energy_share=0.429206
L02: contrast_relative=0.009845, common_relative=0.021091, contrast_energy_share=0.508803
L09: contrast_relative=0.009781, common_relative=0.030301, contrast_energy_share=0.543125

Elapsed seconds: 2.64

Interpretation limits:
Zero-sum head contrasts cancel algebraically under identical head routing.
Whether real prompts produce identical or different routing is unknown.
Common/contrast energy fractions are not causal importance measures.
They do not identify guesses, correct answers, or which checkpoint is sycophantic.
No attention weights, hidden states, nonlinearities, or model inputs were evaluated.

```