=== HYPOTHESIS: MORE LITERAL RESIDUAL-FEATURE COPYING ===
All operations are weight products, SVDs, or eigendecompositions.
No input, activation, normalization denominator, or attention score.
C = learned_post_norm_diagonal @ O_head @ V @ learned_input_norm_diagonal.
These are value-transport coefficient numerators, not full attention.
A and B are compared against the SAME midpoint readable-space projector.
Higher positive-projector fraction means more positive identity-like transport.
Same-space fraction alone does not imply copying: rotations also remain in a subspace.
The control removes the strongest midpoint output axis on both sides of C.

=== AGGREGATE DIRECTIONAL RESULTS ===
All changes are B minus A. Heads are NOT independent observations.
condition metric                 mean_A       mean_B       mean_change    median_change   B>A_heads
raw       alignment                0.0260465    0.0260552   8.695355e-06    -1.082743e-06        34/72
raw       positive_fraction        0.0050999    0.0051030   3.146742e-06                0        20/72
raw       same_space               0.1267136    0.1267159   2.262878e-06    -6.146291e-06        29/72
def       alignment                0.0269838    0.0269812  -2.550054e-06    -2.532126e-06        29/72
def       positive_fraction        0.0073490    0.0073495   4.764742e-07                0        19/72
def       same_space               0.1645027    0.1644974  -5.338484e-06    -3.010518e-06        31/72

=== PER-LAYER MEAN COPYING ALIGNMENT ===
layer raw_A raw_B raw_change def_A def_B def_change raw_fraction_change def_fraction_change
    0 -0.007836 -0.007834  2.2267e-06 -0.015114 -0.015120 -6.50035e-06         3.40334e-07         4.49678e-08
    1  0.007726  0.007717 -9.02685e-06  0.015257  0.015258 1.61566e-06        -2.45566e-07         6.40618e-08
    2 -0.006268 -0.006273 -4.87462e-06 -0.010454 -0.010449 4.23978e-06          3.3758e-08         3.40992e-09
    3  0.011282  0.011273 -8.62862e-06  0.011034  0.011016 -1.8201e-05        -3.15314e-06        -8.97611e-07
    4 -0.001283 -0.001307 -2.41357e-05 -0.001575 -0.001617 -4.14682e-05        -1.93386e-07        -5.22205e-07
    5  0.003051  0.003043 -7.67888e-06  0.005429  0.005400 -2.82199e-05        -1.56989e-07        -7.15516e-07
    6  0.061006  0.060997 -9.09817e-06  0.079149  0.079130 -1.88358e-05        -6.04842e-06        -3.96574e-06
    7  0.049075  0.049195 0.000119672  0.059703  0.059718 1.47468e-05         4.20094e-05        -5.69417e-06
    8  0.058135  0.058198  6.2276e-05  0.041090  0.041138 4.73246e-05         5.49907e-06         1.25021e-05
    9  0.046578  0.046618 3.97271e-05  0.051649  0.051657 8.23131e-06         1.90718e-05         9.50808e-06
   10  0.047648  0.047657 9.16883e-06  0.048662  0.048667 5.16286e-06         1.13833e-06         6.48937e-07
   11  0.020519  0.020510 -8.46483e-06  0.020788  0.020779 -8.81677e-06        -1.76993e-06        -1.88145e-06
   12  0.005075  0.005075 3.35451e-07  0.005168  0.005169 4.42353e-07        -1.72058e-08        -1.98056e-08
   13  0.025515  0.025512 -2.86267e-06  0.025649  0.025646 -3.1306e-06         1.17863e-08        -1.82576e-08
   14  0.034397  0.034393 -3.20466e-06  0.034212  0.034208 -3.6009e-06        -3.64757e-07         -4.0009e-07
   15  0.034939  0.034952 1.23794e-05  0.034805  0.034818 1.34646e-05         2.26565e-06         1.89464e-06
   16  0.033494  0.033487 -6.08907e-06  0.034239  0.034232 -6.90819e-06        -5.53862e-07        -6.47726e-07
   17  0.045784  0.045779 -5.2049e-06  0.046017  0.046011 -5.44719e-06         -1.2256e-06         -1.3271e-06

=== HEADS WITH LARGEST ABSOLUTE COPYING-ALIGNMENT CHANGES ===
layer head alignment_A alignment_B change positive_fraction_change controlled_change same_space_change
    7    3    0.114765    0.115053 0.000287728              6.61248e-05       5.95033e-05       0.000450215
    7    0    0.191244    0.191530 0.000286003              0.000109474      -6.92293e-05       0.000313365
    8    0   -0.176225   -0.176037 0.000187898                        0       6.14062e-05      -0.000128056
    9    2    0.240822    0.241004 0.000181757              8.75751e-05       6.51247e-05       0.000384624
    3    0   -0.040147   -0.040073 7.32495e-05                        0      -3.14477e-05      -7.85064e-05
    9    3    0.107983    0.107910 -7.31018e-05             -1.57821e-05       3.47393e-06      -0.000302841
    3    2    0.076732    0.076662 -7.01885e-05             -1.07664e-05      -1.83519e-05       -4.3402e-05
    7    1   -0.236174   -0.236239 -6.51447e-05                        0       7.81788e-05       5.46381e-05
    6    1   -0.100867   -0.100810 5.71952e-05                        0      -1.53556e-05      -5.79043e-05
    6    2    0.150839    0.150783 -5.68308e-05             -1.71414e-05        1.6157e-06      -4.21392e-05
   15    2    0.091924    0.091974 4.94871e-05               9.1006e-06       4.81151e-05      -1.17858e-05
    8    2    0.212814    0.212857 4.26297e-05              1.81462e-05       0.000116776       1.53143e-05
   11    1    0.084842    0.084799 -4.2239e-05             -7.16546e-06      -4.43511e-05       5.05857e-06
    3    1    0.021935    0.021893 -4.21212e-05             -1.84611e-06      -5.45537e-06      -8.72447e-05
    4    1    0.010514    0.010478 -3.68495e-05             -7.73544e-07      -6.64828e-05      -9.08671e-05
    9    0    0.062641    0.062677 3.58632e-05              4.49432e-06       1.16722e-05       0.000261179
    4    2   -0.002975   -0.003006 -3.04774e-05                        0      -6.16339e-05      -3.29681e-05
    7    2    0.126466    0.126436 -2.98991e-05             -7.56155e-06      -9.46555e-06      -0.000100766
    6    3    0.094234    0.094208 -2.55164e-05             -4.80835e-06      -2.20928e-05      -7.92692e-05
    0    0    0.030183    0.030208  2.4574e-05              1.48404e-06       4.68246e-06       8.57611e-05

Layer-level direction counts:
raw B higher: 7 A higher: 11 ties: 0
def B higher: 8 A higher: 10 ties: 0
Readable-space ranks: [(256, 256)]

Interpretation boundary:
A consistent directional result would support a difference in literal
feature-copying geometry, not establish sycophancy.
The weights tested here do not say which token supplies the copied
information, whether it represents an answer, or whether that answer is true.
