=== EXACT ATTENTION OV-CHANGE DECOMPOSITION ===
Signed inner-product shares: output projection, value projection, post-attn gain, input gain.
Shares sum to 1; negative shares indicate cancellation. This is not behavioral attribution.
layer head | ||delta_OV|| | output value post_gain input_gain | reconstruction_relative_error
15  1 |      1636.5 | +0.364425 +0.635575 +0.000000 +0.000000 | 4.14e-16
15  2 |      1373.9 | +0.531738 +0.468262 +0.000000 +0.000000 | 2.21e-16
15  3 |      1354.6 | +0.593434 +0.406566 +0.000000 +0.000000 | 2.99e-16
12  1 |        1241 | +0.511297 +0.488703 +0.000000 +0.000000 | 5.95e-16
12  2 |      1159.4 | +0.635631 +0.364369 +0.000000 +0.000000 | 4.22e-16
15  0 |      1139.6 | +0.699319 +0.300681 +0.000000 +0.000000 | 4.67e-16
12  3 |       945.3 | +0.679843 +0.320157 +0.000000 -0.000000 | 5.38e-16
14  3 |      876.01 | +0.528383 +0.471617 +0.000000 +0.000000 | 6.11e-16
12  0 |      847.98 | +0.672240 +0.327760 +0.000000 -0.000000 | 3.82e-16
14  2 |      793.92 | +0.558350 +0.441650 +0.000000 +0.000000 | 7.56e-16
16  2 |       773.2 | +0.646145 +0.353855 +0.000000 +0.000000 | 3.24e-16
16  0 |      766.22 | +0.662673 +0.337327 +0.000000 +0.000000 | 2.9e-16
16  1 |      753.96 | +0.674300 +0.325700 +0.000000 +0.000000 | 1.39e-16
14  1 |      742.59 | +0.528357 +0.471643 +0.000000 +0.000000 | 5.56e-16
16  3 |      725.91 | +0.706339 +0.293661 +0.000000 +0.000000 | 1.64e-16
14  0 |      717.89 | +0.549681 +0.450319 +0.000000 +0.000000 | 5.71e-16
13  3 |      716.78 | +0.480309 +0.519691 +0.000000 +0.000000 | 7.52e-16
17  1 |      535.71 | +0.716483 +0.283517 +0.000000 +0.000000 | 2.26e-16
13  0 |      525.45 | +0.629245 +0.370755 +0.000000 +0.000000 | 3.53e-16
17  2 |      523.68 | +0.631336 +0.368664 +0.000000 +0.000000 | 4.09e-16

=== HEADS WITH LARGEST ABSOLUTE NORM-GAIN ATTRIBUTIONS ===
layer= 0 head=0 delta_norm=3.69553 post_gain_share=+0.000010 input_gain_share=+0.000000
layer= 0 head=3 delta_norm=3.86904 post_gain_share=+0.000007 input_gain_share=+0.000000
layer= 2 head=2 delta_norm=24.8959 post_gain_share=+0.000005 input_gain_share=+0.000000
layer= 2 head=0 delta_norm=26.3815 post_gain_share=+0.000004 input_gain_share=+0.000000
layer= 2 head=1 delta_norm=26.6614 post_gain_share=+0.000004 input_gain_share=+0.000000
layer= 2 head=3 delta_norm=19.7674 post_gain_share=+0.000004 input_gain_share=-0.000000
layer= 0 head=1 delta_norm=3.15211 post_gain_share=+0.000003 input_gain_share=+0.000000
layer= 0 head=2 delta_norm=4.29392 post_gain_share=+0.000003 input_gain_share=+0.000000
layer= 3 head=2 delta_norm=44.3764 post_gain_share=+0.000000 input_gain_share=-0.000000
layer= 1 head=0 delta_norm=39.2589 post_gain_share=+0.000000 input_gain_share=+0.000000
layer= 1 head=1 delta_norm=45.2859 post_gain_share=+0.000000 input_gain_share=+0.000000
layer=11 head=3 delta_norm=386.675 post_gain_share=+0.000000 input_gain_share=+0.000000

=== PER-LAYER ATTENTION ATTRIBUTIONS, ENERGY-WEIGHTED ACROSS HEADS ===
0 [0.6279 0.3721 0.     0.    ]
1 [0.3564 0.6436 0.     0.    ]
2 [0.6595 0.3405 0.     0.    ]
3 [ 0.7685  0.2315  0.     -0.    ]
4 [0.8363 0.1637 0.     0.    ]
5 [0.3226 0.6774 0.     0.    ]
6 [ 0.6897  0.3103  0.     -0.    ]
7 [0.4675 0.5325 0.     0.    ]
8 [ 0.528  0.472  0.    -0.   ]
9 [0.5696 0.4304 0.     0.    ]
10 [ 0.6359  0.3641 -0.      0.    ]
11 [0.5228 0.4772 0.     0.    ]
12 [0.6077 0.3923 0.     0.    ]
13 [0.6106 0.3894 0.     0.    ]
14 [0.5405 0.4595 0.     0.    ]
15 [0.5165 0.4835 0.     0.    ]
16 [0.6713 0.3287 0.     0.    ]
17 [0.6346 0.3654 0.     0.    ]
Maximum attention decomposition error: 9.019625627135806e-16

=== EXACT MLP COEFFICIENT-CHANGE DECOMPOSITION ===
Shares refer to the sum of individual unit coefficient-change energies.
Columns: down, gate, up, third-order interaction. Effective factors include gains when marked.
 0 raw: +0.612034 +0.196518 +0.191448 +0.000000 | gain-weighted: +0.839627 +0.090199 +0.070174 +0.000000
 1 raw: +0.783941 +0.108702 +0.107357 +0.000000 | gain-weighted: +0.901751 +0.051693 +0.046556 +0.000000
 2 raw: +0.708491 +0.151128 +0.140381 +0.000000 | gain-weighted: +0.805239 +0.106025 +0.088736 +0.000000
 3 raw: +0.766731 +0.125022 +0.108247 +0.000000 | gain-weighted: +0.825955 +0.110863 +0.063182 +0.000000
 4 raw: +0.809442 +0.102711 +0.087847 +0.000000 | gain-weighted: +0.901361 +0.055521 +0.043118 +0.000000
 5 raw: +0.567566 +0.238356 +0.194079 +0.000000 | gain-weighted: +0.626078 +0.212134 +0.161788 +0.000000
 6 raw: +0.721166 +0.158134 +0.120700 +0.000000 | gain-weighted: +0.720582 +0.156782 +0.122636 +0.000000
 7 raw: +0.674341 +0.186972 +0.138687 +0.000000 | gain-weighted: +0.608316 +0.224895 +0.166789 +0.000000
 8 raw: +0.775948 +0.128197 +0.095855 +0.000000 | gain-weighted: +0.789622 +0.127084 +0.083295 +0.000000
 9 raw: +0.634069 +0.206664 +0.159267 +0.000000 | gain-weighted: +0.628953 +0.211324 +0.159722 +0.000000
10 raw: +0.774923 +0.127721 +0.097356 +0.000000 | gain-weighted: +0.828520 +0.097580 +0.073900 +0.000000
11 raw: +0.645869 +0.194378 +0.159753 +0.000000 | gain-weighted: +0.634189 +0.199887 +0.165925 +0.000000
12 raw: +0.588408 +0.240131 +0.171461 +0.000000 | gain-weighted: +0.578559 +0.227229 +0.194212 +0.000000
13 raw: +0.750704 +0.132732 +0.116565 +0.000000 | gain-weighted: +0.730809 +0.141260 +0.127931 +0.000000
14 raw: +0.775079 +0.122157 +0.102763 +0.000000 | gain-weighted: +0.771771 +0.124508 +0.103721 +0.000000
15 raw: +0.711721 +0.157236 +0.131042 +0.000000 | gain-weighted: +0.711657 +0.158635 +0.129708 +0.000000
16 raw: +0.679161 +0.175922 +0.144917 +0.000000 | gain-weighted: +0.676276 +0.176724 +0.147000 +0.000000
17 raw: +0.686336 +0.183818 +0.129846 +0.000000 | gain-weighted: +0.669980 +0.191485 +0.138534 +0.000000

=== FACTOR ATTRIBUTIONS FOR LARGE UNIT-COEFFICIENT CHANGES ===
layer= 3 unit=1414 delta_norm=    674.78 shares=+0.020683 +0.900286 +0.079031 -0.000000
layer= 3 unit= 226 delta_norm=    515.93 shares=+0.958556 +0.018737 +0.022707 -0.000000
layer= 3 unit= 986 delta_norm=    449.64 shares=+0.967768 +0.012634 +0.019599 -0.000000
layer= 2 unit=1831 delta_norm=    442.21 shares=+0.986966 +0.000376 +0.012657 +0.000000
layer= 3 unit=  82 delta_norm=    436.77 shares=+0.952804 +0.000363 +0.046833 -0.000000
layer= 5 unit= 584 delta_norm=    332.42 shares=+0.014111 +0.594902 +0.390986 -0.000000
layer=11 unit=1815 delta_norm=    278.48 shares=+0.938464 +0.001609 +0.059927 -0.000000
layer=11 unit= 716 delta_norm=    271.01 shares=+0.960815 +0.012097 +0.027088 -0.000000
layer= 4 unit=1044 delta_norm=     267.1 shares=+0.997390 +0.000772 +0.001838 +0.000000
layer=11 unit=1576 delta_norm=    262.74 shares=+0.873311 +0.014414 +0.112276 +0.000000
layer=10 unit=1363 delta_norm=    250.78 shares=+0.992186 +0.001718 +0.006096 -0.000000
layer=14 unit= 641 delta_norm=    241.78 shares=+0.974067 +0.002602 +0.023331 -0.000000
layer=13 unit=  16 delta_norm=    206.96 shares=+0.936165 -0.001166 +0.065001 +0.000000
layer=11 unit=1892 delta_norm=    205.39 shares=+0.723617 +0.030319 +0.246063 -0.000000
layer= 4 unit=1527 delta_norm=    184.03 shares=+0.981451 +0.002036 +0.016513 +0.000000
layer=10 unit=1564 delta_norm=    178.42 shares=+0.949385 +0.010166 +0.040449 +0.000000
layer= 4 unit=1353 delta_norm=     173.4 shares=+0.990784 +0.002357 +0.006859 +0.000000
layer= 6 unit=1122 delta_norm=    171.72 shares=+0.988177 +0.001346 +0.010477 -0.000000
layer= 4 unit= 280 delta_norm=     169.5 shares=+0.990844 +0.000964 +0.008192 +0.000000
layer=12 unit=1528 delta_norm=    167.27 shares=+0.779667 +0.001449 +0.218884 +0.000000
layer=16 unit= 893 delta_norm=    147.13 shares=+0.922532 +0.006314 +0.071155 +0.000000
layer=10 unit=1529 delta_norm=    144.86 shares=+0.994831 +0.000888 +0.004281 +0.000000
layer= 2 unit= 154 delta_norm=    141.49 shares=+0.935824 +0.041623 +0.022553 -0.000000
layer= 2 unit= 925 delta_norm=    135.26 shares=+0.811782 +0.088852 +0.099367 +0.000000

All decompositions are identities involving weights only.
No attention probabilities, nonlinear gates, activations, or predictions were computed.
