OUTPUT of run 6 of at most 60:
```
=== SUMMED MLP COEFFICIENT-TENSOR DIFFERENCES ===
T = sum_i d_i outer sym(u_i outer g_i), with learned diagonal scales.
This is an algebraic coefficient signature, not the nonlinear MLP.
 L relative_delta cross_unit_ratio negative_contributors top1_signed_share top10_signed_share
 0      0.0168772         1.019798                     1          0.043180           0.101784
 1      0.0253330         1.010922                     0          0.009537           0.045967
 2      0.0175069         1.014284                     1          0.004514           0.031589
 3      0.0224219         1.019028                     1          0.015867           0.076070
 4      0.0249918         1.048773                    38          0.025860           0.101564
 5      0.0164277         1.141135                    13          0.036269           0.087240
 6      0.0098526         1.119770                    15          0.143698           0.182523
 7      0.0093277         1.078104                     5          0.032132           0.083499
 8      0.0197394         1.095012                     4          0.023854           0.083598
 9      0.0166753         1.131079                     1          0.005189           0.035095
10      0.0207160         1.117804                     8          0.073595           0.200434
11      0.0145702         1.103527                     3          0.009811           0.053301
12      0.0149452         1.100623                     0          0.007142           0.035696
13      0.0183564         1.074746                     0          0.011779           0.045049
14      0.0207954         1.045584                     0          0.023217           0.051918
15      0.0186891         1.049123                     0          0.010489           0.033669
16      0.0173907         1.065703                     0          0.015392           0.049781
17      0.0187506         1.187840                     0          0.011177           0.056727

=== TOP COHERENT CONTRIBUTORS PER LAYER ===
Tuple: (unit, signed fraction of total difference energy, individual-energy fraction).
Signed fractions measure algebraic overlap, NOT causal behavioral importance.
L00: [(1405, 0.04318, 0.043441), (1856, 0.012139, 0.01161), (878, 0.009756, 0.009012), (1130, 0.008231, 0.008775), (1352, 0.007404, 0.006564)]; top10 overlap with individual-energy ranking=9; negative total share=-0.000009
L01: [(1725, 0.009537, 0.009643), (2035, 0.008054, 0.008048), (827, 0.004698, 0.004625), (534, 0.003904, 0.003926), (606, 0.003452, 0.003393)]; top10 overlap with individual-energy ranking=9; negative total share=0.000000
L02: [(900, 0.004514, 0.004706), (307, 0.003441, 0.003417), (1831, 0.003363, 0.003354), (882, 0.003117, 0.003064), (1049, 0.003098, 0.003001)]; top10 overlap with individual-energy ranking=9; negative total share=-0.000020
L03: [(146, 0.015867, 0.015713), (1414, 0.013054, 0.013869), (1987, 0.008147, 0.008289), (926, 0.006993, 0.007041), (1635, 0.005974, 0.005847)]; top10 overlap with individual-energy ranking=9; negative total share=-0.000001
L04: [(13, 0.02586, 0.025452), (789, 0.011074, 0.009885), (84, 0.010919, 0.011396), (937, 0.009719, 0.009914), (1804, 0.009628, 0.008392)]; top10 overlap with individual-energy ranking=8; negative total share=-0.001070
L05: [(584, 0.036269, 0.044581), (1267, 0.010336, 0.010994), (1100, 0.006764, 0.007537), (725, 0.005545, 0.006928), (13, 0.005526, 0.006736)]; top10 overlap with individual-energy ranking=9; negative total share=-0.000225
L06: [(1824, 0.143698, 0.157604), (1122, 0.008036, 0.009923), (690, 0.005055, 0.005577), (1946, 0.004857, 0.004742), (1293, 0.003892, 0.003363)]; top10 overlap with individual-energy ranking=8; negative total share=-0.000377
L07: [(1300, 0.032132, 0.029784), (1455, 0.012868, 0.013586), (1214, 0.006597, 0.00424), (1057, 0.005533, 0.005822), (656, 0.004951, 0.005802)]; top10 overlap with individual-energy ranking=8; negative total share=-0.000513
L08: [(188, 0.023854, 0.029452), (1500, 0.013675, 0.014375), (456, 0.007659, 0.008975), (227, 0.007306, 0.007041), (1868, 0.006557, 0.00697)]; top10 overlap with individual-energy ranking=8; negative total share=-0.000135
L09: [(808, 0.005189, 0.005592), (437, 0.004364, 0.003247), (448, 0.004331, 0.004767), (1782, 0.003808, 0.003972), (335, 0.003244, 0.002669)]; top10 overlap with individual-energy ranking=7; negative total share=-0.000022
L10: [(1363, 0.073595, 0.076419), (1564, 0.040635, 0.045823), (1529, 0.031589, 0.03), (900, 0.012965, 0.01415), (1869, 0.011962, 0.01165)]; top10 overlap with individual-energy ranking=9; negative total share=-0.000226
L11: [(716, 0.009811, 0.009251), (1576, 0.008136, 0.007043), (1815, 0.007036, 0.009174), (1892, 0.00507, 0.005358), (662, 0.00408, 0.003483)]; top10 overlap with individual-energy ranking=9; negative total share=-0.000040
L12: [(1528, 0.007142, 0.008018), (1520, 0.003795, 0.003998), (2027, 0.00346, 0.003624), (259, 0.003288, 0.003842), (1584, 0.003225, 0.003104)]; top10 overlap with individual-energy ranking=9; negative total share=0.000000
L13: [(16, 0.011779, 0.012648), (1517, 0.005758, 0.005176), (369, 0.004984, 0.004685), (1707, 0.003572, 0.003735), (1292, 0.00335, 0.003498)]; top10 overlap with individual-energy ranking=9; negative total share=0.000000
L14: [(641, 0.023217, 0.024184), (685, 0.004545, 0.003969), (663, 0.003781, 0.003765), (1767, 0.003615, 0.003611), (392, 0.003295, 0.003651)]; top10 overlap with individual-energy ranking=9; negative total share=0.000000
L15: [(1369, 0.010489, 0.010821), (1226, 0.004989, 0.005249), (651, 0.003559, 0.00339), (1940, 0.002762, 0.00255), (1337, 0.002568, 0.002653)]; top10 overlap with individual-energy ranking=6; negative total share=0.000000
L16: [(893, 0.015392, 0.016547), (1591, 0.007359, 0.008764), (730, 0.007073, 0.007377), (1593, 0.003886, 0.003196), (287, 0.00368, 0.00248)]; top10 overlap with individual-energy ranking=8; negative total share=0.000000
L17: [(1026, 0.011177, 0.004493), (1916, 0.009069, 0.008981), (1650, 0.008251, 0.004281), (1312, 0.008161, 0.007587), (1853, 0.003821, 0.000943)]; top10 overlap with individual-energy ranking=7; negative total share=0.000000

=== RECHECK OF PREVIOUSLY HIGHLIGHTED UNITS ===
L00 unit=1405 coherent_rank=   1 signed_share=+0.0431797 individual_energy/total_difference=0.0425973
L01 unit=1725 coherent_rank=   1 signed_share=+0.0095366 individual_energy/total_difference=0.0095388
L01 unit=2035 coherent_rank=   2 signed_share=+0.0080536 individual_energy/total_difference=0.0079608
L03 unit= 146 coherent_rank=   1 signed_share=+0.0158668 individual_energy/total_difference=0.0154200
L03 unit=1414 coherent_rank=   2 signed_share=+0.0130542 individual_energy/total_difference=0.0136105
L04 unit=  13 coherent_rank=   1 signed_share=+0.0258602 individual_energy/total_difference=0.0242685
L04 unit=  84 coherent_rank=   3 signed_share=+0.0109186 individual_energy/total_difference=0.0108656
L04 unit= 789 coherent_rank=   2 signed_share=+0.0110742 individual_energy/total_difference=0.0094252
L04 unit= 937 coherent_rank=   4 signed_share=+0.0097186 individual_energy/total_difference=0.0094525
L04 unit=1044 coherent_rank=   6 signed_share=+0.0088411 individual_energy/total_difference=0.0091712
L04 unit=1320 coherent_rank=   7 signed_share=+0.0074113 individual_energy/total_difference=0.0068623
L04 unit=1804 coherent_rank=   5 signed_share=+0.0096284 individual_energy/total_difference=0.0080018
L05 unit= 584 coherent_rank=   1 signed_share=+0.0362690 individual_energy/total_difference=0.0390671
L06 unit=1824 coherent_rank=   1 signed_share=+0.1436979 individual_energy/total_difference=0.1407468
L08 unit= 188 coherent_rank=   1 signed_share=+0.0238543 individual_energy/total_difference=0.0268963
L08 unit=1500 coherent_rank=   2 signed_share=+0.0136750 individual_energy/total_difference=0.0131275
L10 unit=1363 coherent_rank=   1 signed_share=+0.0735950 individual_energy/total_difference=0.0683654
L10 unit=1564 coherent_rank=   2 signed_share=+0.0406347 individual_energy/total_difference=0.0409939
L10 unit=1529 coherent_rank=   3 signed_share=+0.0315893 individual_energy/total_difference=0.0268386
L14 unit= 641 coherent_rank=   1 signed_share=+0.0232172 individual_energy/total_difference=0.0231292

=== LAYER RANKING BY SUMMED-COEFFICIENT RELATIVE DIFFERENCE ===
L01: relative=0.0253330 cross_unit_ratio=1.010922
L04: relative=0.0249918 cross_unit_ratio=1.048773
L03: relative=0.0224219 cross_unit_ratio=1.019028
L14: relative=0.0207954 cross_unit_ratio=1.045584
L10: relative=0.0207160 cross_unit_ratio=1.117804
L08: relative=0.0197394 cross_unit_ratio=1.095012
L17: relative=0.0187506 cross_unit_ratio=1.187840
L15: relative=0.0186891 cross_unit_ratio=1.049123
L13: relative=0.0183564 cross_unit_ratio=1.074746
L02: relative=0.0175069 cross_unit_ratio=1.014284
L16: relative=0.0173907 cross_unit_ratio=1.065703
L00: relative=0.0168772 cross_unit_ratio=1.019798
L09: relative=0.0166753 cross_unit_ratio=1.131079
L05: relative=0.0164277 cross_unit_ratio=1.141135
L12: relative=0.0149452 cross_unit_ratio=1.100623
L11: relative=0.0145702 cross_unit_ratio=1.103527
L06: relative=0.0098526 cross_unit_ratio=1.119770
L07: relative=0.0093277 cross_unit_ratio=1.078104

Maximum independent energy-accounting discrepancy: 3.7146017040849614e-11
Elapsed seconds: 8.41

Interpretation:
cross_unit_ratio > 1 means unit changes reinforce in this coefficient tensor;
< 1 means net cancellation. Negative contributors oppose its total change.
Symmetrization retains the shared-input gate/value coefficient but discards
their nonlinear asymmetry. The separate gate analysis remains necessary.
No activation frequency, semantic role, or behavioral effect was measured.

```