=== EXACT EVEN-PART COEFFICIENT AGGREGATION ===
For the linear-gain MLP numerator, the even part is
  (1/2) sum_i d_i (g_i^T z)(u_i^T z).
This expression is used symbolically only; no z or activation is evaluated.
Stored pre/post diagonal gains are included; RMS denominators are not.

=== COMPLETE LAYER RESULTS: ALL 2048 UNITS ===
layer | whole_delta_norm individual_RSS coherence relative_whole_delta | top10_attribution negative_attribution negative_unit_fraction
 2 |      1110.15      1099.64 1.019204 0.0191339 | 0.072021 -0.000300 0.000977
 3 |      3060.04      3017.57 1.028342 0.0231088 | 0.057108 -0.000003 0.000488
10 |      476.895      445.318 1.146846 0.0205425 | 0.197047 -0.000211 0.005371
12 |      938.298      893.957 1.101663 0.0149434 | 0.039206 +0.000000 0.000000
16 |       579.16      560.384 1.068137 0.0178494 | 0.048855 +0.000000 0.000000

Coherence = ||sum_i delta T_i||^2 / sum_i ||delta T_i||^2.
Signed unit attribution = <delta T_i, sum_j delta T_j> / ||sum_j delta T_j||^2.
Attributions sum to one, but are not behavioral or intervention effects.
Maximum diagonal consistency error: 1.0195634148287204e-11
Attribution sums: [0.9999999999999998, 1.0000000000000009, 1.0000000000000002, 1.0000000000000007, 0.9999999999999989]

=== TOP CONTRIBUTORS TO EACH COMPLETE EVEN-PART CHANGE ===
layer unit | signed_attribution individual_energy_share | whole/individual ranks | gate_up_cosine
 2 1831 | +0.0395990 0.0404289 |    1/1    | -0.869917
 2    6 | +0.0043300 0.0028316 |    2/9    | +0.418340
 2  154 | +0.0042421 0.0041392 |    3/2    | +0.076939
 2  918 | +0.0039057 0.0036061 |    4/5    | -0.034977
 2  798 | +0.0037884 0.0024579 |    5/14   | -0.292748
 2  925 | +0.0037192 0.0037827 |    6/3    | +0.073199
 2  700 | +0.0033247 0.0037125 |    7/4    | -0.126367
 2  900 | +0.0032115 0.0035176 |    8/6    | -0.082592
 3 1414 | +0.0125146 0.0125013 |    1/1    | -0.331086
 3  226 | +0.0066545 0.0073081 |    2/2    | +0.268875
 3 1671 | +0.0058986 0.0051025 |    3/5    | -0.900462
 3  986 | +0.0054151 0.0055508 |    4/3    | -0.171416
 3   82 | +0.0050591 0.0052377 |    5/4    | +0.609112
 3   95 | +0.0046663 0.0045802 |    6/6    | +0.031035
 3 1558 | +0.0043980 0.0040200 |    7/8    | -0.396377
 3 1635 | +0.0043220 0.0043418 |    8/7    | +0.325965
10 1363 | +0.0780284 0.0792859 |    1/1    | -0.695510
10 1564 | +0.0346243 0.0401300 |    2/2    | +0.632092
10 1529 | +0.0296607 0.0264527 |    3/3    | -0.713648
10 1106 | +0.0127541 0.0103162 |    4/5    | +0.801652
10  900 | +0.0111535 0.0113768 |    5/4    | -0.836264
10 1869 | +0.0091823 0.0093956 |    6/6    | +0.714293
10 1003 | +0.0059796 0.0062000 |    7/7    | +0.779464
10 1016 | +0.0054765 0.0042037 |    8/12   | -0.095355
12 1528 | +0.0080724 0.0087527 |    1/1    | -0.757138
12 1520 | +0.0040502 0.0048016 |    2/2    | -0.249119
12  259 | +0.0037781 0.0038265 |    3/3    | +0.583161
12   22 | +0.0037572 0.0034956 |    4/5    | -0.343036
12 2027 | +0.0035525 0.0035898 |    5/4    | -0.498739
12    6 | +0.0032724 0.0031570 |    6/8    | +0.599859
12 1584 | +0.0032555 0.0031139 |    7/9    | -0.701235
12  644 | +0.0032203 0.0028280 |    8/13   | -0.516979
16  893 | +0.0156962 0.0172325 |    1/1    | +0.786388
16  730 | +0.0071423 0.0076187 |    2/2    | +0.520316
16 1591 | +0.0069032 0.0076078 |    3/3    | +0.120783
16  287 | +0.0044650 0.0030751 |    4/4    | -0.338827
16  324 | +0.0032879 0.0026788 |    5/6    | -0.514582
16 1446 | +0.0028652 0.0021003 |    6/8    | -0.196571
16  375 | +0.0023135 0.0009667 |    7/80   | -0.081664
16  238 | +0.0022171 0.0021295 |    8/7    | -0.071553

=== PREVIOUSLY HIGHLIGHTED UNITS ===
layer= 2 unit=1831 whole_rank=   1 individual_rank=   1 signed_attribution=+0.0395990 gate_up_cosine=-0.869917
layer= 3 unit=  82 whole_rank=   5 individual_rank=   4 signed_attribution=+0.0050591 gate_up_cosine=+0.609112
layer= 3 unit= 226 whole_rank=   2 individual_rank=   2 signed_attribution=+0.0066545 gate_up_cosine=+0.268875
layer= 3 unit= 986 whole_rank=   4 individual_rank=   3 signed_attribution=+0.0054151 gate_up_cosine=-0.171416
layer= 3 unit=1414 whole_rank=   1 individual_rank=   1 signed_attribution=+0.0125146 gate_up_cosine=-0.331086
layer=10 unit=1363 whole_rank=   1 individual_rank=   1 signed_attribution=+0.0780284 gate_up_cosine=-0.695510
layer=12 unit= 259 whole_rank=   3 individual_rank=   3 signed_attribution=+0.0037781 gate_up_cosine=+0.583161
layer=12 unit=1520 whole_rank=   2 individual_rank=   2 signed_attribution=+0.0040502 gate_up_cosine=-0.249119
layer=12 unit=1528 whole_rank=   1 individual_rank=   1 signed_attribution=+0.0080724 gate_up_cosine=-0.757138
layer=16 unit= 893 whole_rank=   1 individual_rank=   1 signed_attribution=+0.0156962 gate_up_cosine=+0.786388

=== STRONGEST CANCELLING UNIT CHANGES IN EACH LAYER ===
layer= 2 unit= 246 signed_attribution=-0.0001958 individual_energy_share=0.0010635
layer= 2 unit= 649 signed_attribution=-0.0001040 individual_energy_share=0.0003437
layer= 2 unit=1825 signed_attribution=+0.0000179 individual_energy_share=0.0000195
layer= 3 unit=1831 signed_attribution=-0.0000034 individual_energy_share=0.0000156
layer= 3 unit= 523 signed_attribution=+0.0000036 individual_energy_share=0.0000030
layer= 3 unit=1949 signed_attribution=+0.0000084 individual_energy_share=0.0000160
layer=10 unit= 241 signed_attribution=-0.0000758 individual_energy_share=0.0006458
layer=10 unit=1474 signed_attribution=-0.0000347 individual_energy_share=0.0001033
layer=10 unit=  33 signed_attribution=-0.0000311 individual_energy_share=0.0001017
layer=12 unit= 881 signed_attribution=+0.0000207 individual_energy_share=0.0000878
layer=12 unit= 495 signed_attribution=+0.0000294 individual_energy_share=0.0000968
layer=12 unit=1212 signed_attribution=+0.0000340 individual_energy_share=0.0000553
layer=16 unit=2016 signed_attribution=+0.0000301 individual_energy_share=0.0000805
layer=16 unit= 405 signed_attribution=+0.0000780 individual_energy_share=0.0001243
layer=16 unit=  23 signed_attribution=+0.0000815 individual_energy_share=0.0001842

Interpretation limits:
- The even part is exact for the specified unnormalized numerator, not the whole normalized MLP block.
- The odd, nonlinear part can differ independently and is not characterized by this tensor.
- A coefficient Frobenius metric does not specify the distribution of actual hidden states.
- Unit attribution here concerns an algebraic sum, not semantic responsibility or causal necessity.
- No inputs, activations, probabilities, model execution, or training were used.
