OUTPUT of run 5 of at most 60:
```
ADJACENT WRITER–READER COMPOSITION TEST
All products are between weight matrices, not model inputs.
Includes learned diagonal scales; omits state-dependent denominators,
attention routing, activation functions, and intervening nonlinear effects.
These are algebraic numerator connections, NOT complete network pathways.
Interference < 1: compensation; > 1: reinforcement; = 1: orthogonal changes.
Reader-energy share excludes the interference term.

Number of connections: 87
Direct-product validation relative error: 2.697340833914228e-16

ALL CONNECTIONS count = 87
  relative                     min/median/max: [0.0094 0.0179 0.0465]
  interference                 min/median/max: [1.0049 1.0241 1.085 ]
  term_cosine                  min/median/max: [0.0054 0.0286 0.1273]
  reader_energy_share          min/median/max: [0.0546 0.2155 0.5543]
  norm_gain_B_minus_A          min/median/max: [-0.0004  0.      0.0003]
  coupling_gain_B_minus_A      min/median/max: [-0.0003  0.      0.0006]
  Compensation count: 0
  Reinforcement count: 87

attention_to_gate count = 18
  relative                     min/median/max: [0.0094 0.0132 0.0179]
  interference                 min/median/max: [1.0119 1.0249 1.044 ]
  term_cosine                  min/median/max: [0.0125 0.027  0.0529]
  reader_energy_share          min/median/max: [0.1548 0.2897 0.4157]
  norm_gain_B_minus_A          min/median/max: [-0.0001  0.      0.0001]
  coupling_gain_B_minus_A      min/median/max: [-0.      0.      0.0002]
  Compensation count: 0
  Reinforcement count: 18

attention_to_up count = 18
  relative                     min/median/max: [0.0116 0.014  0.0181]
  interference                 min/median/max: [1.0115 1.0191 1.0404]
  term_cosine                  min/median/max: [0.0116 0.021  0.049 ]
  reader_energy_share          min/median/max: [0.1638 0.3094 0.5543]
  norm_gain_B_minus_A          min/median/max: [-0.      0.      0.0001]
  coupling_gain_B_minus_A      min/median/max: [-0.0001  0.      0.0001]
  Compensation count: 0
  Reinforcement count: 18

mlp_to_k count = 17
  relative                     min/median/max: [0.0153 0.0228 0.0465]
  interference                 min/median/max: [1.0137 1.0274 1.0474]
  term_cosine                  min/median/max: [0.0204 0.0384 0.0736]
  reader_energy_share          min/median/max: [0.0546 0.1656 0.3728]
  norm_gain_B_minus_A          min/median/max: [-0.0004 -0.0001  0.0002]
  coupling_gain_B_minus_A      min/median/max: [-0.0003  0.      0.0005]
  Compensation count: 0
  Reinforcement count: 17

mlp_to_q count = 17
  relative                     min/median/max: [0.016  0.024  0.0321]
  interference                 min/median/max: [1.0049 1.0138 1.0199]
  term_cosine                  min/median/max: [0.0054 0.0166 0.0244]
  reader_energy_share          min/median/max: [0.1346 0.227  0.401 ]
  norm_gain_B_minus_A          min/median/max: [-0.      0.      0.0003]
  coupling_gain_B_minus_A      min/median/max: [-0.0002  0.      0.0006]
  Compensation count: 0
  Reinforcement count: 17

mlp_to_v count = 17
  relative                     min/median/max: [0.0141 0.0195 0.0287]
  interference                 min/median/max: [1.0235 1.0483 1.085 ]
  term_cosine                  min/median/max: [0.0336 0.0753 0.1273]
  reader_energy_share          min/median/max: [0.0575 0.136  0.2108]
  norm_gain_B_minus_A          min/median/max: [-0.0001 -0.      0.0002]
  coupling_gain_B_minus_A      min/median/max: [-0.0001  0.      0.0004]
  Compensation count: 0
  Reinforcement count: 17

STRONGEST COMPENSATION
L13.mlp -> L14.attention.q               rel=0.0253404 interference=1.00492 term_cos=0.00541611 reader_share=29.132% norm_gain_B-A=4.96896e-05 coupling_gain_B-A=0.000134875
L10.mlp -> L11.attention.q               rel=0.0246937 interference=1.00859 term_cos=0.0121206 reader_share=14.711% norm_gain_B-A=-4.71387e-05 coupling_gain_B-A=-0.000127457
L08.mlp -> L09.attention.q               rel=0.0239714 interference=1.00943 term_cos=0.0118719 reader_share=19.624% norm_gain_B-A=1.26814e-05 coupling_gain_B-A=0.000159036
L07.mlp -> L08.attention.q               rel=0.0196997 interference=1.00975 term_cos=0.0114744 reader_share=23.655% norm_gain_B-A=1.30236e-05 coupling_gain_B-A=-5.29071e-05
L15.mlp -> L16.attention.q               rel=0.028801 interference=1.01085 term_cos=0.0157169 reader_share=13.824% norm_gain_B-A=4.09766e-05 coupling_gain_B-A=0.000615965
L06.mlp -> L07.attention.q               rel=0.0210083 interference=1.0109 term_cos=0.0130118 reader_share=22.700% norm_gain_B-A=-1.45778e-05 coupling_gain_B-A=5.76176e-05
L00.attention -> L00.mlp.up              rel=0.0125899 interference=1.0115 term_cos=0.0115638 reader_share=55.430% norm_gain_B-A=-6.25226e-06 coupling_gain_B-A=0.000124772
L00.attention -> L00.mlp.gate            rel=0.00938968 interference=1.01187 term_cos=0.0125332 reader_share=33.945% norm_gain_B-A=1.65098e-06 coupling_gain_B-A=6.14511e-05

STRONGEST REINFORCEMENT
L05.mlp -> L06.attention.v               rel=0.0149335 interference=1.08498 term_cos=0.106581 reader_share=19.820% norm_gain_B-A=-2.15131e-05 coupling_gain_B-A=4.37609e-05
L03.mlp -> L04.attention.v               rel=0.0204391 interference=1.08176 term_cos=0.127297 reader_share=11.675% norm_gain_B-A=-6.2379e-05 coupling_gain_B-A=0.000179302
L09.mlp -> L10.attention.v               rel=0.0185066 interference=1.078 term_cos=0.100096 reader_share=18.662% norm_gain_B-A=3.30734e-05 coupling_gain_B-A=-4.42712e-05
L11.mlp -> L12.attention.v               rel=0.019531 interference=1.06249 term_cos=0.0803292 reader_share=18.580% norm_gain_B-A=3.6813e-05 coupling_gain_B-A=7.14022e-05
L02.mlp -> L03.attention.v               rel=0.0171534 interference=1.06113 term_cos=0.0924468 reader_share=12.491% norm_gain_B-A=-6.72457e-05 coupling_gain_B-A=0.000364563
L10.mlp -> L11.attention.v               rel=0.028744 interference=1.05191 term_cos=0.0721678 reader_share=15.265% norm_gain_B-A=0.000175253 coupling_gain_B-A=0.000196796
L15.mlp -> L16.attention.v               rel=0.0219515 interference=1.05162 term_cos=0.0753023 reader_share=13.596% norm_gain_B-A=-5.02086e-05 coupling_gain_B-A=0.00014689
L01.mlp -> L02.attention.v               rel=0.020986 interference=1.05002 term_cos=0.0840814 reader_share=9.810% norm_gain_B-A=-3.49599e-05 coupling_gain_B-A=1.50616e-05

LARGEST RELATIVE COMPOSITION DIFFERENCES
L04.mlp -> L05.attention.k               rel=0.0465488 interference=1.03163 term_cos=0.0696215 reader_share=5.458% norm_gain_B-A=-0.000437211 coupling_gain_B-A=7.84812e-05
L10.mlp -> L11.attention.k               rel=0.0365352 interference=1.03251 term_cos=0.0629713 reader_share=7.179% norm_gain_B-A=2.40438e-05 coupling_gain_B-A=-6.74581e-05
L16.mlp -> L17.attention.q               rel=0.0321027 interference=1.01416 term_cos=0.0154855 reader_share=29.770% norm_gain_B-A=0.000260953 coupling_gain_B-A=-0.000244072
L04.mlp -> L05.attention.q               rel=0.0319847 interference=1.01275 term_cos=0.016567 reader_share=18.087% norm_gain_B-A=0.000296693 coupling_gain_B-A=0.000193515
L14.mlp -> L15.attention.q               rel=0.031658 interference=1.01292 term_cos=0.0188812 reader_share=13.533% norm_gain_B-A=-1.98516e-05 coupling_gain_B-A=-0.000155602
L15.mlp -> L16.attention.k               rel=0.0306381 interference=1.02794 term_cos=0.0460222 reader_share=10.269% norm_gain_B-A=0.000231245 coupling_gain_B-A=0.000329681
L14.mlp -> L15.attention.k               rel=0.029473 interference=1.0223 term_cos=0.0384387 reader_share=9.279% norm_gain_B-A=-0.000325329 coupling_gain_B-A=0.000537752
L11.mlp -> L12.attention.k               rel=0.0292749 interference=1.04536 term_cos=0.0736331 reader_share=10.613% norm_gain_B-A=-0.000185273 coupling_gain_B-A=0.00017474
L15.mlp -> L16.attention.q               rel=0.028801 interference=1.01085 term_cos=0.0157169 reader_share=13.824% norm_gain_B-A=4.09766e-05 coupling_gain_B-A=0.000615965
L10.mlp -> L11.attention.v               rel=0.028744 interference=1.05191 term_cos=0.0721678 reader_share=15.265% norm_gain_B-A=0.000175253 coupling_gain_B-A=0.000196796

Interpretation boundary:
Compensation would show that some separate parameter changes overstate
changes in these composed matrices; reinforcement would show the reverse.
Signed B-minus-A gains identify algebraic gain changes, not increased
sycophancy, copying of a user's guess, or correctness.

```