=== EARLIER-MLP TO LATER-GATE COEFFICIENT CONNECTIONS ===
All source<target pairs are examined through the direct residual-skip route.
Coupling gain compares coefficient orientation with a uniformly rotated
orientation calibration; it is not an activation or influence estimate.
Completed source MLP 00
Completed source MLP 01
Completed source MLP 02
Completed source MLP 03
Completed source MLP 04
Completed source MLP 05
Completed source MLP 06
Completed source MLP 07
Completed source MLP 08
Completed source MLP 09
Completed source MLP 10
Completed source MLP 11
Completed source MLP 12
Completed source MLP 13
Completed source MLP 14
Completed source MLP 15
Completed source MLP 16

=== HIGHLIGHTED GATES: STRONGEST PRECEDING MLP CONNECTIONS ===
A gain of 1 is the orientation calibration, not a behavioral threshold.
Source/control compares the changed coefficient orientation with independent
output-coordinate sign flips preserving its spectrum and diagonal energies.
target unit degree baseline_source baseline_gain change_source change_gain source/control
     4  937      2               0      0.013680             0    0.009309       1.008844
     4  937      3               0      0.013976             0    0.009876       1.048506
     4 1804      2               0      0.014772             0    0.007321       0.752578
     4 1804      3               3      0.010085             0    0.007306       0.736610
     4 1044      2               0      0.017759             0    0.011859       1.083628
     4 1044      3               0      0.012139             0    0.011036       0.975740
     6 1824      2               0      0.055036             0    0.057827       1.181615
     6 1824      3               5      0.042131             5    0.053226       1.094826
    10 1363      2               0      0.049661             0    0.021943       1.213315
    10 1363      3               0      0.041938             0    0.023374       1.312008
    10 1564      2               0      0.050941             0    0.045444       0.979023
    10 1564      3               0      0.066578             0    0.045337       1.029684
    10 1529      2               0      0.045815             0    0.021975       1.349692
    10 1529      3               0      0.037609             0    0.022141       1.369176
    14  641      2              10      0.141657             0    0.021627       1.216576
    14  641      3              10      0.199720             0    0.021956       1.237104
    16  893      2              10      0.152164             0    0.027602       1.178106
    16  893      3              14      0.552465             0    0.027082       1.099790
    16 1593      2               0      0.152644            15    0.392046      14.294258
    16 1593      3               1      0.145199            15    0.442620      15.996273

=== TOP THREE SOURCE CONNECTIONS FOR LAYER-10 CANDIDATES ===
unit=1363, degree=2: (source,baseline_gain,change_gain,source/control)=[(0, 0.049661, 0.021943, 1.213315), (2, 0.011946, 0.010151, 0.940573), (1, 0.017617, 0.008821, 0.921314)]
unit=1363, degree=3: (source,baseline_gain,change_gain,source/control)=[(0, 0.041938, 0.023374, 1.312008), (7, 0.021226, 0.017213, 3.695814), (2, 0.012343, 0.011974, 1.076397)]
unit=1564, degree=2: (source,baseline_gain,change_gain,source/control)=[(0, 0.050941, 0.045444, 0.979023), (9, 0.028757, 0.029987, 1.104044), (3, 0.031896, 0.028277, 1.146126)]
unit=1564, degree=3: (source,baseline_gain,change_gain,source/control)=[(0, 0.066578, 0.045337, 1.029684), (9, 0.029847, 0.034033, 1.202487), (2, 0.029175, 0.030506, 1.03645)]
unit=1529, degree=2: (source,baseline_gain,change_gain,source/control)=[(0, 0.045815, 0.021975, 1.349692), (2, 0.010331, 0.00931, 1.004633), (1, 0.016223, 0.008067, 0.970977)]
unit=1529, degree=3: (source,baseline_gain,change_gain,source/control)=[(0, 0.037609, 0.022141, 1.369176), (7, 0.018278, 0.016007, 3.369662), (2, 0.009687, 0.009922, 1.03067)]

=== WHOLE-LAYER CONTEXT ===
Per target, median over gates of the strongest earlier-source change gain.
These maxima are exploratory selections, not significance tests.
L01: degree2: median=0.02700, top(unit,source,gain)=[(1007, 0, 0.0956), (951, 0, 0.08), (998, 0, 0.0784)] | degree3: median=0.02712, top(unit,source,gain)=[(1987, 0, 0.092), (1626, 0, 0.087), (1007, 0, 0.0861)]
L02: degree2: median=0.01945, top(unit,source,gain)=[(1924, 1, 0.0689), (1278, 1, 0.0677), (0, 1, 0.0668)] | degree3: median=0.01980, top(unit,source,gain)=[(1924, 1, 0.0676), (1278, 1, 0.0634), (0, 1, 0.0623)]
L03: degree2: median=0.01731, top(unit,source,gain)=[(45, 0, 0.0831), (877, 2, 0.0591), (754, 2, 0.0584)] | degree3: median=0.01811, top(unit,source,gain)=[(45, 0, 0.0888), (1471, 0, 0.0622), (692, 0, 0.0588)]
L04: degree2: median=0.01521, top(unit,source,gain)=[(2037, 0, 0.0863), (143, 0, 0.0714), (2035, 0, 0.0664)] | degree3: median=0.01573, top(unit,source,gain)=[(143, 0, 0.075), (1518, 0, 0.0727), (2037, 0, 0.0712)]
L05: degree2: median=0.02319, top(unit,source,gain)=[(1021, 0, 0.4585), (1734, 0, 0.1438), (1851, 0, 0.1399)] | degree3: median=0.02258, top(unit,source,gain)=[(1021, 0, 0.5417), (1734, 0, 0.1648), (1851, 0, 0.1459)]
L06: degree2: median=0.03603, top(unit,source,gain)=[(1142, 1, 0.3962), (2005, 0, 0.3925), (1317, 1, 0.3797)] | degree3: median=0.03865, top(unit,source,gain)=[(2005, 0, 0.4381), (1142, 1, 0.3786), (1317, 1, 0.3639)]
L07: degree2: median=0.05716, top(unit,source,gain)=[(210, 0, 0.4366), (455, 0, 0.2979), (1203, 0, 0.2558)] | degree3: median=0.03922, top(unit,source,gain)=[(210, 0, 0.5041), (155, 0, 0.3078), (455, 0, 0.2875)]
L08: degree2: median=0.12291, top(unit,source,gain)=[(706, 7, 0.5124), (2019, 0, 0.4617), (65, 0, 0.3842)] | degree3: median=0.10158, top(unit,source,gain)=[(706, 7, 0.8091), (188, 7, 0.7435), (1425, 7, 0.6793)]
L09: degree2: median=0.12362, top(unit,source,gain)=[(256, 0, 0.6245), (166, 0, 0.6023), (173, 0, 0.5496)] | degree3: median=0.12872, top(unit,source,gain)=[(256, 0, 0.5587), (166, 0, 0.5236), (1379, 7, 0.4939)]
L10: degree2: median=0.10804, top(unit,source,gain)=[(179, 0, 0.964), (1093, 0, 0.8509), (566, 0, 0.7422)] | degree3: median=0.11200, top(unit,source,gain)=[(179, 0, 0.7975), (1003, 0, 0.7879), (1745, 0, 0.7834)]
L11: degree2: median=0.11768, top(unit,source,gain)=[(558, 1, 3.0631), (1575, 0, 1.3632), (886, 0, 1.2457)] | degree3: median=0.11841, top(unit,source,gain)=[(558, 1, 2.9316), (1575, 0, 1.6048), (235, 0, 1.1479)]
L12: degree2: median=0.16175, top(unit,source,gain)=[(1435, 0, 1.6508), (595, 0, 1.5672), (1187, 0, 1.445)] | degree3: median=0.16472, top(unit,source,gain)=[(1435, 0, 1.4791), (1667, 0, 1.4132), (649, 0, 1.4024)]
L13: degree2: median=0.12762, top(unit,source,gain)=[(360, 1, 2.129), (30, 0, 1.6275), (633, 0, 1.3505)] | degree3: median=0.12993, top(unit,source,gain)=[(360, 1, 2.0355), (30, 0, 1.6044), (633, 0, 1.3304)]
L14: degree2: median=0.12406, top(unit,source,gain)=[(1676, 1, 2.7868), (490, 0, 1.0288), (769, 0, 1.0047)] | degree3: median=0.12412, top(unit,source,gain)=[(1676, 1, 2.665), (490, 0, 1.0639), (796, 0, 0.9982)]
L15: degree2: median=0.12202, top(unit,source,gain)=[(1921, 1, 2.4914), (1200, 0, 1.8339), (713, 0, 1.2696)] | degree3: median=0.12317, top(unit,source,gain)=[(1921, 1, 2.386), (1200, 0, 1.7833), (713, 0, 1.3129)]
L16: degree2: median=0.12424, top(unit,source,gain)=[(425, 0, 2.9787), (614, 0, 2.2903), (853, 1, 1.8711)] | degree3: median=0.12415, top(unit,source,gain)=[(425, 0, 2.9676), (614, 0, 2.0759), (853, 1, 1.7854)]
L17: degree2: median=0.13772, top(unit,source,gain)=[(737, 0, 2.1976), (1268, 0, 1.8434), (1079, 0, 1.627)] | degree3: median=0.13891, top(unit,source,gain)=[(1268, 0, 2.1153), (737, 0, 2.028), (1079, 0, 1.5624)]

Minimum coefficient-covariance eigenvalue / largest eigenvalue: -1.6052269588412437e-20
Elapsed seconds: 17.66

Interpretation limits:
Only direct residual-skip coefficient connections are examined.
Intervening nonlinear processing, state-dependent normalization, routing,
and cancellation with other branches are not evaluated.
Quadratic and cubic coefficient orientations do not specify activation
frequencies or the importance of either order on real residual states.
These connections cannot by themselves identify guesses, truth, or sycophancy.
