OUTPUT of run 3 of at most 60:
```
=== ATTENTION: LEARNED-DIAGONAL COMPOSED OPERATORS ===
Q*=diag(1+q_norm) Q_h diag(1+input_norm)
K*=diag(1+k_norm) K   diag(1+input_norm)
O*=diag(1+post_attention_norm) O_h
V*=V diag(1+input_norm)
Comparing Q*.T K* and O* V*.
These omit input-dependent denominators, RoPE, and attention weights.

Per layer: QK_relative[H0..H3], OV_relative[H0..H3]
L00 QK 0.01218 0.00917 0.00797 0.01040  OV 0.00923 0.02170 0.00742 0.00764
L01 QK 0.00949 0.00975 0.01267 0.00975  OV 0.01086 0.01065 0.00860 0.00835
L02 QK 0.01091 0.00876 0.01207 0.01443  OV 0.01226 0.01223 0.01028 0.02012
L03 QK 0.00867 0.01889 0.01259 0.00653  OV 0.01513 0.01994 0.01508 0.02208
L04 QK 0.01227 0.01738 0.01472 0.01318  OV 0.02117 0.03144 0.02140 0.01363
L05 QK 0.02297 0.01414 0.01836 0.02461  OV 0.00759 0.01502 0.01444 0.01183
L06 QK 0.01513 0.01763 0.02319 0.01629  OV 0.01748 0.01186 0.01711 0.01398
L07 QK 0.01252 0.01673 0.01457 0.02052  OV 0.01471 0.00981 0.01113 0.01826
L08 QK 0.01271 0.01170 0.01638 0.01356  OV 0.01303 0.01741 0.01995 0.01512
L09 QK 0.01855 0.01390 0.01868 0.01245  OV 0.01381 0.01109 0.01415 0.01347
L10 QK 0.02253 0.03276 0.01036 0.01782  OV 0.01090 0.01199 0.01493 0.00763
L11 QK 0.04660 0.03806 0.01812 0.02737  OV 0.01229 0.02081 0.01020 0.00856
L12 QK 0.03024 0.02385 0.03112 0.02583  OV 0.01413 0.01290 0.00896 0.01272
L13 QK 0.01689 0.02174 0.02365 0.01200  OV 0.01051 0.01292 0.01355 0.00834
L14 QK 0.01740 0.02587 0.01792 0.02001  OV 0.00821 0.00931 0.01051 0.00882
L15 QK 0.02271 0.01460 0.01881 0.01336  OV 0.01215 0.01664 0.01885 0.01527
L16 QK 0.01464 0.01664 0.01602 0.01167  OV 0.01258 0.01436 0.01215 0.01427
L17 QK 0.01554 0.02298 0.00890 0.01488  OV 0.00690 0.01463 0.00908 0.01088

Top 16 by combined relative discrepancy:
layer head QK_rel OV_rel QK_angle OV_angle QK_delta OV_delta
   11    0 0.046603 0.012294 0.046529 0.012293 37.785 253.715
   11    1 0.038063 0.020808 0.038063 0.020808 44.7313 233.797
    4    1 0.017382 0.031444 0.017381 0.031438 667.864 86.9788
   10    1 0.032763 0.011986 0.032762 0.011986 312.333 209.281
   12    0 0.030236 0.014131 0.030194 0.014130 383.573 870.736
   12    2 0.031121 0.008959 0.031076 0.008958 387.696 1079.37
    6    2 0.023186 0.017112 0.023186 0.017107 263.006 60.9628
   12    3 0.025833 0.012720 0.025831 0.012714 342.776 935.818
   11    3 0.027371 0.008560 0.027371 0.008560 34.0211 400.373
   14    1 0.025865 0.009308 0.025840 0.009307 626.032 738.128
    7    3 0.020517 0.018261 0.020516 0.018254 315.938 74.7313
    3    1 0.018886 0.019943 0.018886 0.019859 320.225 44.6888
    5    3 0.024607 0.011830 0.024606 0.011823 383.282 94.8868
   13    2 0.023653 0.013550 0.023653 0.013543 231.452 428.697
   17    1 0.022978 0.014634 0.022975 0.014633 383.101 499.817
   12    1 0.023847 0.012895 0.023840 0.012894 363.461 1260.37

=== MLP UNIT COUPLINGS ===
For unit j, use its gate row g, up row u, and down column d.
P_j = d outer u: exact gate-conditioned rank-one linear coefficient.
T_j = d outer sym(g outer u): algebraic coupling of all three factors.
T is NOT a substitute for the nonlinear MLP.
Norms of these objects are computed by contractions, without constructing inputs.
For scaled results: g,u include the pre-MLP diagonal; d includes the post-MLP diagonal.
Layer summaries sum UNIT energies; they do not include cross-unit interference.

Layer summaries:
layer raw_P_rel raw_T_rel scaled_P_rel scaled_T_rel scaled_T_delta top10_T_delta_energy%
    0 0.012017 0.012746 0.017478 0.016573 50.9517 10.168
    1 0.017824 0.018741 0.024315 0.025224 1855.19 4.597
    2 0.014949 0.016161 0.016260 0.017391 2017.28 3.171
    3 0.016814 0.017817 0.020799 0.022481 5882.12 7.724
    4 0.019230 0.020445 0.027469 0.024469 1882.65 10.131
    5 0.011769 0.013751 0.014870 0.015701 1829.38 9.985
    6 0.015374 0.017369 0.014970 0.009288 779.15 19.863
    7 0.014092 0.016279 0.012630 0.008936 595.193 8.029
    8 0.017425 0.019415 0.016304 0.019052 495.978 9.314
    9 0.013318 0.015626 0.013032 0.015471 283.035 3.555
   10 0.017242 0.019255 0.017474 0.019689 910.176 21.012
   11 0.013380 0.015194 0.012251 0.013817 3063.35 5.571
   12 0.012659 0.015257 0.012199 0.014121 1789.02 3.845
   13 0.017492 0.019110 0.016237 0.017555 1520.12 4.620
   14 0.019260 0.021118 0.018434 0.020164 1510.65 5.235
   15 0.016528 0.018644 0.016179 0.018074 1007.91 3.361
   16 0.014859 0.017133 0.014716 0.016660 1093.14 4.988
   17 0.014438 0.017379 0.014486 0.017272 1143.68 4.488

Ranked by raw_t_delta
layer unit raw_P_delta raw_T_delta scaled_P_delta scaled_T_delta scaled_P_rel scaled_T_rel scaled_gate_rel
    6 1824 0.011606 0.0356 10.753 309.32 0.00420 0.00421 0.00047
    7 1455 0.0085748 0.031289 2.1869 69.375 0.00127 0.00127 0.00023
    0 1130 0.0099864 0.028008 0.28407 4.7729 0.00469 0.00483 0.00373
    9  237 0.0067108 0.019034 1.3814 11.783 0.03115 0.03113 0.00276
    8  456 0.0072371 0.018506 3.1676 46.988 0.03806 0.03814 0.00194
   10 1564 0.007497 0.01823 6.5097 194.84 0.01358 0.01365 0.00136
   10  488 0.010381 0.018153 5.1567 66.476 0.03201 0.03204 0.00213
    4  563 0.0093832 0.017639 5.3084 158.98 0.04713 0.04718 0.00244
    6 1122 0.0083154 0.017295 3.5377 77.614 0.03234 0.03244 0.00202
    0 1856 0.0087117 0.017069 0.46294 5.4901 0.01791 0.01808 0.00217
   14 1767 0.0071829 0.016827 7.8655 90.778 0.03379 0.03389 0.00231
    8 1406 0.0067092 0.016673 1.2182 11.794 0.03939 0.03938 0.00312

Ranked by scaled_t_delta
layer unit raw_P_delta raw_T_delta scaled_P_delta scaled_T_delta scaled_P_rel scaled_T_rel scaled_gate_rel
    3  146 0.010195 0.010591 18.489 737.34 0.02373 0.02394 0.00334
    3 1414 0.008653 0.0043232 17.748 692.73 0.00438 0.01464 0.01381
    3 1987 0.0064833 0.0076496 11.136 535.53 0.02128 0.02131 0.00263
    3  926 0.0066951 0.0087555 15.926 493.57 0.05380 0.05390 0.00483
    3 1426 0.0072788 0.007331 16.001 476.56 0.02574 0.02619 0.00519
    3 1635 0.0067132 0.0079294 9.6516 449.79 0.00524 0.00677 0.00444
    3 1877 0.0088457 0.0081407 17.06 442.13 0.04424 0.04517 0.00814
    3 1928 0.0071927 0.008404 10.719 431.08 0.01701 0.01713 0.00388
    3  303 0.0034589 0.0034926 5.9298 407.24 0.03249 0.03223 0.00196
    5  584 0.0083914 0.007469 17.891 386.26 0.00460 0.00761 0.00605
    3 1671 0.0039273 0.0042012 5.6969 376.65 0.02504 0.02437 0.00232
    3  126 0.0070048 0.0085436 11.802 375.97 0.02923 0.02946 0.00392
    3 1334 0.0080229 0.0076018 14.148 375.65 0.04491 0.04648 0.00739
    3 1151 0.0067313 0.011411 8.1838 358.94 0.01384 0.01450 0.00418
    3 1247 0.0065463 0.0049073 14.861 349.38 0.03800 0.03863 0.00654
    3   38 0.0062529 0.0077426 10.059 331.56 0.04093 0.04106 0.00418
    3   82 0.0066488 0.010187 3.9951 331.01 0.02875 0.02816 0.00217
    3 1381 0.0065742 0.0056125 12.315 322.97 0.03608 0.03651 0.00552

Overlap of raw and scaled top-100 unit couplings: 4

All calculations above are weight products or contractions.
They establish neither activation frequency nor behavioural causality.

```