=== ATTENTION PRODUCT ATTRIBUTION ===
Shares are signed inner-product allocations of operator-difference energy.
Q+K shares and O+V shares each sum to 1; they are NOT behavioural attributions.
Learned diagonals included; state denominators, RoPE and softmax excluded.
layer head QK_rel Q_share% K_share% OV_rel O_share% V_share% QK_crosscos OV_crosscos
    1    2 0.012666 65.04 34.96 0.008597 36.66 63.34 0.0924 0.1654
    3    3 0.006525 57.59 42.41 0.022078 77.52 22.48 0.0780 0.0869
    4    0 0.012270 47.43 52.57 0.021169 82.67 17.33 0.0235 0.1230
    4    1 0.017382 53.23 46.77 0.031444 87.37 12.63 0.0396 0.1665
    5    2 0.018362 86.48 13.52 0.014438 40.10 59.90 0.0532 0.1084
    6    2 0.023186 41.90 58.10 0.017112 70.08 29.92 0.0615 0.1789
    8    2 0.016382 44.23 55.77 0.019953 87.51 12.49 0.0561 0.0167
   10    1 0.032763 59.01 40.99 0.011986 70.39 29.61 0.0592 0.1443
   11    0 0.046603 31.13 68.87 0.012294 84.29 15.71 0.0494 0.0981
   11    1 0.038063 21.88 78.12 0.020808 89.36 10.64 0.0348 0.1367
   12    0 0.030236 77.15 22.85 0.014131 66.74 33.26 0.0935 0.1967
   12    1 0.023847 66.61 33.39 0.012895 51.92 48.08 0.0776 0.3021
   12    2 0.031121 78.31 21.69 0.008959 60.39 39.61 0.0364 0.2694
   16    1 0.016644 54.37 45.63 0.014362 69.14 30.86 0.0525 0.0793
   16    3 0.011669 64.31 35.69 0.014269 70.75 29.25 0.0729 0.0822
   17    3 0.014885 80.81 19.19 0.010881 75.95 24.05 0.0528 0.0325

=== EXACT MLP GATE-CONDITIONED COEFFICIENT DIFFERENCES ===
For each unit P_j = d_j outer u_j.
Delta P_j = Delta d_j outer u_mid,j + d_mid,j outer Delta u_j.
The gate row is assessed separately; its nonlinear response is never evaluated.
Scaled d/u/g include the learned post/pre-MLP diagonals.
All aggregate P energies here sum per-unit energies, not cross-unit terms.
layer P_rel gate_rel down_share% up_share% top10_P_energy%
    0 0.017478 0.006208 90.404 9.596 4.051
    1 0.024315 0.006348 93.709 6.291 3.495
    2 0.016260 0.006372 86.566 13.434 2.856
    3 0.020799 0.007492 92.496 7.504 4.544
    4 0.027469 0.006155 95.381 4.619 4.989
    5 0.014870 0.006421 85.392 14.608 4.704
    6 0.014970 0.006448 87.144 12.856 4.647
    7 0.012630 0.006220 79.268 20.732 3.849
    8 0.016304 0.006370 89.014 10.986 3.748
    9 0.013032 0.006375 78.009 21.991 2.275
   10 0.017474 0.006267 88.170 11.830 1.905
   11 0.012251 0.006418 77.124 22.876 1.587
   12 0.012199 0.006287 72.882 27.118 1.812
   13 0.016237 0.006480 84.749 15.251 1.163
   14 0.018434 0.006721 87.832 12.168 1.193
   15 0.016179 0.006661 84.100 15.900 1.330
   16 0.014716 0.006537 81.745 18.255 1.520
   17 0.014486 0.006618 81.419 18.581 1.627

Top 20 units by fraction of THEIR LAYER'S coefficient-difference energy:
This removes cross-layer scale differences, but does not establish importance.
layer unit layer_share% P_delta P_rel gate_rel gate_angle down_share%
    0 1405 1.2251 1.21404 0.004043 0.005363 0.005166 1.155
    5  584 1.2219 17.8911 0.004601 0.006049 0.006027 0.948
    6 1824 1.1327 10.7531 0.004198 0.000471 0.000471 97.473
    3 1256 0.9670 29.0192 0.038022 0.024057 0.024055 99.196
    4   84 0.7343 12.434 0.006157 0.012182 0.011898 0.599
    9  437 0.6769 5.6892 0.010102 0.007446 0.007430 43.825
    4  789 0.6579 11.769 0.048993 0.004817 0.004816 99.548
    7 1451 0.6443 7.0971 0.054498 0.010064 0.009948 99.088
    8  188 0.5902 6.3093 0.002641 0.008295 0.008206 11.500
    4 1189 0.5488 10.7485 0.106925 0.009708 0.009703 99.931
    3 1932 0.5204 21.2893 0.030808 0.018450 0.018439 98.762
    8 1917 0.5169 5.90472 0.085277 0.010378 0.010355 99.180
    4  162 0.5053 10.3138 0.133252 0.006841 0.006831 100.279
    4 1761 0.4884 10.1402 0.072549 0.008213 0.008211 99.052
    5 1451 0.4826 11.2435 0.055499 0.011861 0.011799 99.251
    1  534 0.4715 5.56821 0.078235 0.007455 0.007454 99.130
    1  827 0.4662 5.53683 0.103627 0.005330 0.005325 99.668
    7 1452 0.4661 6.03609 0.068380 0.010793 0.010767 99.299
    4   47 0.4545 9.78211 0.044177 0.004546 0.004511 98.242
    7 1119 0.4543 5.95916 0.020262 0.008303 0.008051 91.292

=== CONCENTRATION OF COHERENT MLP PARAMETER DIFFERENCES ===
Each projection is truncated algebraically to ranks 4,8,16 by SVD.
Unit energies combine gate rows, up rows and down columns.
This diagnoses sparse versus distributed UPDATE SUPPORT, not sparse causal support.
layer rank8_top10_energy% top8_rank8_units rank4/rank16_top20_overlap
    0 1.938 [1433, 1683, 949, 95, 513, 1790, 1306, 892] 9/20
    1 1.870 [204, 110, 1113, 698, 1933, 135, 1523, 459] 11/20
    2 1.943 [767, 0, 1597, 862, 1013, 1455, 1806, 1306] 14/20
    3 1.661 [90, 408, 380, 1340, 1334, 1891, 1877, 669] 10/20
    4 1.489 [1204, 1818, 73, 1848, 1792, 1248, 1028, 1935] 14/20
    5 1.575 [57, 660, 1676, 799, 643, 538, 1181, 1158] 15/20
    6 1.451 [929, 187, 1378, 1668, 1146, 1375, 1385, 1727] 17/20
    7 2.162 [256, 186, 1334, 1525, 1875, 1539, 1397, 1957] 17/20
    8 1.627 [706, 329, 599, 707, 787, 1687, 1807, 618] 15/20
    9 1.630 [647, 1476, 1146, 46, 1116, 1950, 380, 1759] 15/20
   10 1.599 [404, 1958, 539, 15, 803, 43, 546, 568] 15/20
   11 1.490 [1723, 1258, 1565, 1282, 1533, 215, 2008, 743] 14/20
   12 1.394 [1961, 1887, 1018, 543, 645, 1424, 504, 252] 15/20
   13 1.371 [954, 1242, 1313, 984, 1756, 687, 1073, 20] 16/20
   14 1.403 [1163, 1293, 1430, 786, 557, 559, 890, 912] 13/20
   15 1.222 [1599, 652, 1207, 842, 1049, 538, 1360, 497] 9/20
   16 1.219 [931, 238, 1724, 876, 1415, 598, 381, 1125] 12/20
   17 1.289 [956, 580, 2044, 1149, 1620, 768, 1037, 390] 14/20
