OUTPUT of run 4 of at most 60:
```
=== MLP UNIT-FACTOR DIFFERENCES BY LAYER ===
Statistics sum INDIVIDUAL unit signature energies.
They do not equal the energy of the summed MLP computation.
 L raw_rel scaled_rel top1_share top10_share effective_units raw/scaled_top10_overlap
 0  0.01299    0.01744    0.03212     0.08752          451.87                        4
 1  0.01875    0.02533    0.00927     0.04553          929.88                        3
 2  0.01618    0.01744    0.00472     0.03091         1225.58                        0
 3  0.01784    0.02260    0.01466     0.07299          675.73                        1
 4  0.02042    0.02443    0.02867     0.10025          375.52                        3
 5  0.01380    0.01570    0.04572     0.10073          298.72                        3
 6  0.01737    0.00985    0.13434     0.17484           52.73                        3
 7  0.01632    0.00913    0.01910     0.06793          673.57                        1
 8  0.01945    0.01901    0.02964     0.09194          457.04                        4
 9  0.01569    0.01554    0.00490     0.03449         1214.34                        2
10  0.01932    0.01961    0.05770     0.16260          159.60                        5
11  0.01528    0.01388    0.00574     0.04544          924.34                        3
12  0.01534    0.01420    0.00581     0.03335         1221.78                        1
13  0.01914    0.01762    0.01018     0.04104         1191.07                        7
14  0.02115    0.02015    0.01601     0.04153         1093.83                        2
15  0.01866    0.01810    0.00925     0.03135         1425.87                        5
16  0.01714    0.01671    0.01152     0.04350         1201.06                        4
17  0.01738    0.01734    0.00576     0.03612         1383.26                        0

=== DECOMPOSITION OF SCALED SIGNATURE DIFFERENCES ===
Percentages sum to 100 including cross terms; cross terms may be negative.
 L write% value% gate% triple% cross%
 0   81.2248    8.4199   10.1869    0.0000    0.1684
 1   88.2333    5.5696    6.0979    0.0000    0.0992
 2   75.3186   11.3847   13.1466    0.0000    0.1501
 3   80.7208    7.0334   12.1693    0.0000    0.0765
 4   89.2399    5.0576    5.7624    0.0000   -0.0600
 5   68.9187   12.8866   17.9965    0.0000    0.1982
 6   76.4981   10.1296   13.5702    0.0000   -0.1979
 7   62.9694   15.7308   21.1361    0.0000    0.1637
 8   76.2612    9.0495   14.4383    0.0000    0.2510
 9   62.5946   15.7199   21.3016    0.0000    0.3839
10   81.5140    8.0042   10.3861    0.0000    0.0957
11   61.4476   16.9809   21.4612    0.0000    0.1103
12   56.6686   19.8082   23.4110    0.0000    0.1121
13   71.7642   13.1900   14.8839    0.0000    0.1619
14   76.6359   10.4268   12.7978    0.0000    0.1396
15   69.8129   13.3811   16.6020    0.0000    0.2041
16   66.6258   14.9553   18.1973    0.0000    0.2216
17   64.7216   14.5808   20.4324    0.0000    0.2652

=== TOP THREE UNITS PER LAYER ===
Each tuple: (unit, within-layer signature-difference share, gate angle in degrees)
L00: [(1405, 0.032122, 0.29598), (1856, 0.011425, 0.12438), (878, 0.00805, 0.15927)]; raw top three=[1130, 1856, 878]
L01: [(1725, 0.009274, 0.17453), (2035, 0.007981, 0.28816), (827, 0.004644, 0.30509)]; raw top three=[1725, 2035, 1528]
L02: [(900, 0.004723, 0.29694), (307, 0.003444, 0.40385), (882, 0.003086, 0.28764)]; raw top three=[421, 1831, 624]
L03: [(146, 0.014659, 0.19115), (1414, 0.012558, 0.79027), (1987, 0.007292, 0.15083)]; raw top three=[1151, 146, 82]
L04: [(13, 0.028669, 0.23296), (84, 0.011789, 0.68173), (789, 0.009987, 0.27596)]; raw top three=[563, 13, 1044]
L05: [(584, 0.045722, 0.34532), (1267, 0.011156, 0.18207), (1100, 0.007623, 0.26453)]; raw top three=[501, 1663, 1034]
L06: [(1824, 0.134344, 0.027), (1122, 0.009167, 0.11562), (690, 0.005274, 0.13164)]; raw top three=[1824, 1122, 1429]
L07: [(1300, 0.019098, 0.05497), (1455, 0.013207, 0.01296), (1057, 0.006003, 0.25421)]; raw top three=[1455, 1214, 825]
L08: [(188, 0.029642, 0.47018), (1500, 0.013515, 0.13045), (456, 0.009177, 0.11068)]; raw top three=[456, 1406, 1377]
L09: [(448, 0.004898, 0.11521), (808, 0.003986, 0.1978), (437, 0.003983, 0.42571)]; raw top three=[237, 448, 957]
L10: [(1363, 0.0577, 0.05194), (1564, 0.036735, 0.07668), (1529, 0.022351, 0.0647)]; raw top three=[488, 1564, 1301]
L11: [(716, 0.005736, 0.16518), (1815, 0.005571, 0.07397), (1576, 0.005508, 0.14399)]; raw top three=[1388, 1892, 575]
L12: [(1528, 0.005808, 0.06608), (1520, 0.003859, 0.12662), (244, 0.003392, 0.17036)]; raw top three=[715, 1823, 1590]
L13: [(16, 0.010184, 0.09768), (1517, 0.004861, 0.11832), (369, 0.004297, 0.08484)]; raw top three=[859, 1575, 369]
L14: [(641, 0.016005, 0.05934), (1767, 0.003749, 0.13222), (392, 0.003704, 0.15494)]; raw top three=[1767, 809, 416]
L15: [(1369, 0.009246, 0.04777), (1226, 0.005602, 0.30723), (651, 0.003113, 0.0698)]; raw top three=[651, 1554, 1369]
L16: [(893, 0.011516, 0.05145), (1591, 0.009422, 0.39354), (730, 0.006123, 0.10632)]; raw top three=[1589, 1829, 1672]
L17: [(1312, 0.005765, 0.09878), (1916, 0.005621, 0.11214), (1650, 0.004571, 0.39026)]; raw top three=[1217, 546, 824]

=== TWENTY LEADING UNIT CANDIDATES UNDER THE DECLARED PROXY ===
Score = unit difference energy / sum of baseline unit-signature energies in its layer.
This ranks a weight-factor proxy, NOT demonstrated behavioral effects.
L10 unit=1363 score=2.219197e-05 share=0.05770 T_rel=0.02961 DU_rel=0.02960 gate_angle_deg=0.05194 gate_norm_B/A=1.000025
L04 unit=  13 score=1.711464e-05 share=0.02867 T_rel=0.00642 DU_rel=0.00502 gate_angle_deg=0.23296 gate_norm_B/A=0.998891
L10 unit=1564 score=1.412844e-05 share=0.03673 T_rel=0.01367 DU_rel=0.01358 gate_angle_deg=0.07668 gate_norm_B/A=0.999778
L06 unit=1824 score=1.303845e-05 share=0.13434 T_rel=0.00423 DU_rel=0.00420 gate_angle_deg=0.02700 gate_norm_B/A=0.999997
L05 unit= 584 score=1.126374e-05 share=0.04572 T_rel=0.00763 DU_rel=0.00460 gate_angle_deg=0.34532 gate_norm_B/A=1.000515
L08 unit= 188 score=1.070886e-05 share=0.02964 T_rel=0.00872 DU_rel=0.00264 gate_angle_deg=0.47018 gate_norm_B/A=0.998790
L00 unit=1405 score=9.770603e-06 share=0.03212 T_rel=0.00691 DU_rel=0.00404 gate_angle_deg=0.29598 gate_norm_B/A=0.998561
L10 unit=1529 score=8.596475e-06 share=0.02235 T_rel=0.03430 DU_rel=0.03429 gate_angle_deg=0.06470 gate_norm_B/A=0.999918
L03 unit= 146 score=7.48804e-06 share=0.01466 T_rel=0.02398 DU_rel=0.02373 gate_angle_deg=0.19115 gate_norm_B/A=0.999825
L04 unit=  84 score=7.037603e-06 share=0.01179 T_rel=0.01353 DU_rel=0.00616 gate_angle_deg=0.68173 gate_norm_B/A=0.997391
L14 unit= 641 score=6.499076e-06 share=0.01601 T_rel=0.02479 DU_rel=0.02479 gate_angle_deg=0.05934 gate_norm_B/A=0.999860
L03 unit=1414 score=6.415016e-06 share=0.01256 T_rel=0.01449 DU_rel=0.00438 gate_angle_deg=0.79027 gate_norm_B/A=0.999224
L04 unit= 789 score=5.962226e-06 share=0.00999 T_rel=0.04926 DU_rel=0.04899 gate_angle_deg=0.27596 gate_norm_B/A=1.000029
L01 unit=1725 score=5.948349e-06 share=0.00927 T_rel=0.05895 DU_rel=0.05894 gate_angle_deg=0.17453 gate_norm_B/A=1.000158
L01 unit=2035 score=5.118666e-06 share=0.00798 T_rel=0.06287 DU_rel=0.06271 gate_angle_deg=0.28816 gate_norm_B/A=1.000187
L04 unit=1044 score=5.013431e-06 share=0.00840 T_rel=0.05538 DU_rel=0.05532 gate_angle_deg=0.15346 gate_norm_B/A=1.000169
L08 unit=1500 score=4.882406e-06 share=0.01351 T_rel=0.03020 DU_rel=0.03005 gate_angle_deg=0.13045 gate_norm_B/A=0.999911
L04 unit= 937 score=4.791198e-06 share=0.00803 T_rel=0.13512 DU_rel=0.13538 gate_angle_deg=0.14219 gate_norm_B/A=0.999687
L04 unit=1320 score=4.422297e-06 share=0.00741 T_rel=0.07228 DU_rel=0.07216 gate_angle_deg=0.24374 gate_norm_B/A=0.999976
L04 unit=1804 score=4.182315e-06 share=0.00701 T_rel=0.20573 DU_rel=0.20582 gate_angle_deg=0.11258 gate_norm_B/A=1.000137

=== GATE-DIRECTION AND MAGNITUDE SUMMARY ===
L00: gate angle median/p99/max=0.35170/0.64153/1.04494 deg; gate norm ratio median/min/max=0.999975/0.998013/1.002854
L01: gate angle median/p99/max=0.35758/0.59816/0.81856 deg; gate norm ratio median/min/max=0.999998/0.998755/1.002230
L02: gate angle median/p99/max=0.35925/0.74342/1.06495 deg; gate norm ratio median/min/max=0.999986/0.997896/1.001982
L03: gate angle median/p99/max=0.45056/1.22318/2.03926 deg; gate norm ratio median/min/max=0.999983/0.996916/1.003508
L04: gate angle median/p99/max=0.41731/1.06707/1.65052 deg; gate norm ratio median/min/max=0.999998/0.996762/1.003464
L05: gate angle median/p99/max=0.42219/1.31255/1.67872 deg; gate norm ratio median/min/max=1.000002/0.997100/1.003587
L06: gate angle median/p99/max=0.42580/0.88334/1.27009 deg; gate norm ratio median/min/max=1.000018/0.997145/1.002400
L07: gate angle median/p99/max=0.42564/0.80835/1.03878 deg; gate norm ratio median/min/max=1.000007/0.997203/1.003850
L08: gate angle median/p99/max=0.44642/0.87316/1.53312 deg; gate norm ratio median/min/max=1.000003/0.996751/1.003491
L09: gate angle median/p99/max=0.43488/0.99928/1.38120 deg; gate norm ratio median/min/max=1.000002/0.994034/1.004852
L10: gate angle median/p99/max=0.43085/1.02663/1.33238 deg; gate norm ratio median/min/max=1.000012/0.994333/1.005447
L11: gate angle median/p99/max=0.42552/1.00029/1.70580 deg; gate norm ratio median/min/max=0.999993/0.995108/1.006495
L12: gate angle median/p99/max=0.42163/0.86419/1.18948 deg; gate norm ratio median/min/max=1.000016/0.995006/1.003903
L13: gate angle median/p99/max=0.40388/0.70573/1.31017 deg; gate norm ratio median/min/max=0.999993/0.995735/1.004185
L14: gate angle median/p99/max=0.41786/0.70283/0.98871 deg; gate norm ratio median/min/max=1.000029/0.994951/1.005665
L15: gate angle median/p99/max=0.42288/0.65478/0.74385 deg; gate norm ratio median/min/max=0.999981/0.995060/1.003981
L16: gate angle median/p99/max=0.41184/0.66685/0.85485 deg; gate norm ratio median/min/max=1.000033/0.995895/1.004427
L17: gate angle median/p99/max=0.44806/0.74072/1.39182 deg; gate norm ratio median/min/max=0.999974/0.994925/1.004239

Interpretation boundary:
The DU outer product removes exact reciprocal up/down rescaling ambiguity.
The three-factor signature describes coupled gate/read/write weights,
but is not a substitute for GELU, activation frequencies, normalization
denominators, inter-unit cancellation, or downstream decoding.
No semantic label or causal importance follows from these signatures alone.

```