=== ADJACENT MLP-TO-ATTENTION COEFFICIENT LINKS ===
Readers are Frobenius-normalized weight matrices, not normalized states.
No intervening nonlinear operation or attention routing is evaluated.
src->dst sharedK_rel strongestQ Q_rel  sharedK_coupling_gain
00->01    0.015390         Q0 0.017187                0.04153
01->02    0.024029         Q3 0.025710                0.02151
02->03    0.020998         Q3 0.024231                0.00851
03->04    0.025985         Q3 0.031164                0.00444
04->05    0.044713         Q1 0.036148                0.00125
05->06    0.022797         Q1 0.022685                0.03318
06->07    0.024104         Q3 0.025224                0.01379
07->08    0.022725         Q3 0.026243                0.02518
08->09    0.030871         Q0 0.028383                0.08771
09->10    0.027684         Q0 0.027542                0.08236
10->11    0.047877         Q2 0.034774                0.03588
11->12    0.032897         Q3 0.032283                0.01640
12->13    0.024510         Q3 0.030140                0.01002
13->14    0.027938         Q2 0.030493                0.01893
14->15    0.031538         Q1 0.034484                0.01589
15->16    0.033401         Q2 0.031927                0.01760
16->17    0.023457         Q3 0.037133                0.00956

=== TOP 20 LINKS BY RELATIVE COEFFICIENT DIFFERENCE ===
Source/receiver/cross fractions refer to algebraic coefficient changes.
They are NOT causal behavioral attributions.
MLP10->L11/Kshared relative=0.047877 coupling=0.036 source/receiver/cross=91.42/5.08/3.51% top_units=[1363, 1529, 1564]
MLP04->L05/Kshared relative=0.044713 coupling=0.001 source/receiver/cross=91.37/5.43/3.20% top_units=[13, 1044, 240]
MLP16->L17/Q3      relative=0.037133 coupling=0.016 source/receiver/cross=71.40/23.27/5.33% top_units=[893, 1123, 89]
MLP16->L17/Q0      relative=0.036927 coupling=0.017 source/receiver/cross=75.84/23.27/0.90% top_units=[893, 89, 1593]
MLP04->L05/Q1      relative=0.036148 coupling=0.008 source/receiver/cross=72.11/26.61/1.28% top_units=[13, 1044, 563]
MLP10->L11/Q2      relative=0.034774 coupling=0.071 source/receiver/cross=79.41/18.44/2.14% top_units=[1363, 1529, 1564]
MLP14->L15/Q1      relative=0.034484 coupling=0.021 source/receiver/cross=84.89/14.56/0.55% top_units=[641, 1767, 685]
MLP14->L15/Q3      relative=0.034338 coupling=0.021 source/receiver/cross=83.60/12.78/3.62% top_units=[641, 1503, 663]
MLP14->L15/Q2      relative=0.033861 coupling=0.024 source/receiver/cross=88.09/9.97/1.94% top_units=[641, 685, 663]
MLP15->L16/Kshared relative=0.033401 coupling=0.018 source/receiver/cross=88.71/8.41/2.88% top_units=[1369, 685, 1497]
MLP14->L15/Q0      relative=0.033121 coupling=0.023 source/receiver/cross=88.15/10.10/1.75% top_units=[641, 1286, 1767]
MLP04->L05/Q0      relative=0.032905 coupling=0.009 source/receiver/cross=87.76/11.62/0.61% top_units=[1044, 13, 563]
MLP11->L12/Kshared relative=0.032897 coupling=0.016 source/receiver/cross=83.79/10.86/5.35% top_units=[716, 1576, 881]
MLP10->L11/Q3      relative=0.032546 coupling=0.061 source/receiver/cross=87.91/12.36/-0.27% top_units=[1363, 1529, 1564]
MLP04->L05/Q3      relative=0.032355 coupling=0.013 source/receiver/cross=82.37/16.80/0.83% top_units=[13, 280, 1823]
MLP16->L17/Q1      relative=0.032323 coupling=0.019 source/receiver/cross=75.91/21.69/2.39% top_units=[893, 89, 730]
MLP16->L17/Q2      relative=0.032285 coupling=0.020 source/receiver/cross=62.04/38.25/-0.29% top_units=[893, 407, 1801]
MLP11->L12/Q3      relative=0.032283 coupling=0.021 source/receiver/cross=67.09/30.94/1.97% top_units=[1815, 165, 494]
MLP11->L12/Q1      relative=0.032146 coupling=0.019 source/receiver/cross=74.04/24.99/0.96% top_units=[1892, 494, 1815]
MLP11->L12/Q0      relative=0.031952 coupling=0.023 source/receiver/cross=64.33/32.26/3.41% top_units=[165, 1815, 433]

=== TOP 12 LINKS BY BASELINE COUPLING GAIN ===
Strong baseline coupling need not imply a large or important behavioral change.
MLP09->L10/Q0      coupling_gain=0.1171 relative_difference=0.027542
MLP09->L10/Q3      coupling_gain=0.1048 relative_difference=0.027407
MLP09->L10/Q1      coupling_gain=0.1030 relative_difference=0.024097
MLP09->L10/Q2      coupling_gain=0.1005 relative_difference=0.025874
MLP08->L09/Q3      coupling_gain=0.0991 relative_difference=0.028250
MLP08->L09/Q2      coupling_gain=0.0963 relative_difference=0.027853
MLP08->L09/Q0      coupling_gain=0.0956 relative_difference=0.028383
MLP08->L09/Q1      coupling_gain=0.0890 relative_difference=0.028004
MLP08->L09/Kshared coupling_gain=0.0877 relative_difference=0.030871
MLP09->L10/Kshared coupling_gain=0.0824 relative_difference=0.027684
MLP00->L01/Q2      coupling_gain=0.0733 relative_difference=0.016358
MLP10->L11/Q2      coupling_gain=0.0710 relative_difference=0.034774

=== FOLLOW-UP OF PREVIOUSLY HIGHLIGHTED MLP UNITS ===
MLP06->L07/Q0      (unit, coherent_rank, signed_share)=[(1824, 1, 0.029189)]
MLP06->L07/Q1      (unit, coherent_rank, signed_share)=[(1824, 1, 0.031429)]
MLP06->L07/Q2      (unit, coherent_rank, signed_share)=[(1824, 1, 0.031806)]
MLP06->L07/Q3      (unit, coherent_rank, signed_share)=[(1824, 1, 0.035444)]
MLP06->L07/Kshared (unit, coherent_rank, signed_share)=[(1824, 1, 0.058359)]
MLP10->L11/Q0      (unit, coherent_rank, signed_share)=[(1363, 1, 0.088725), (1564, 2, 0.051832), (1529, 3, 0.046645)]
MLP10->L11/Q1      (unit, coherent_rank, signed_share)=[(1363, 1, 0.083713), (1564, 2, 0.044867), (1529, 3, 0.029406)]
MLP10->L11/Q2      (unit, coherent_rank, signed_share)=[(1363, 1, 0.106944), (1564, 3, 0.02991), (1529, 2, 0.03544)]
MLP10->L11/Q3      (unit, coherent_rank, signed_share)=[(1363, 1, 0.078962), (1564, 3, 0.03395), (1529, 2, 0.04057)]
MLP10->L11/Kshared (unit, coherent_rank, signed_share)=[(1363, 1, 0.078991), (1564, 3, 0.049352), (1529, 2, 0.050656)]
MLP14->L15/Q0      (unit, coherent_rank, signed_share)=[(641, 1, 0.024157)]
MLP14->L15/Q1      (unit, coherent_rank, signed_share)=[(641, 1, 0.022499)]
MLP14->L15/Q2      (unit, coherent_rank, signed_share)=[(641, 1, 0.023764)]
MLP14->L15/Q3      (unit, coherent_rank, signed_share)=[(641, 1, 0.025991)]
MLP14->L15/Kshared (unit, coherent_rank, signed_share)=[(641, 1, 0.014355)]

Maximum energy-accounting discrepancy: 6.992176947821116e-12
Elapsed seconds: 27.84

Interpretation boundary:
These contractions expose potential direct read/write connections.
They do not establish that the relevant units activate, that attention
selects a guessed answer, or that the connection affects an output token.
The omitted nonlinearities and unknown residual-state distribution prevent
converting this coefficient-link ranking into a behavioral importance ranking.
