OUTPUT of run 5 of at most 60:
```
=== MAGNITUDE-MATCHED RANDOMIZED-MATRIX CONTROLS ===
Four independent sign randomizations per selected difference matrix.
Controls preserve entry magnitudes, hence row/column norm anisotropy.
They are not a model of the actual fine-tuning process.
layer part observed_s1 null_s1_mean null_s1_max observed/null_max observed_top4% null_top4_mean%
   12 q     0.105941 0.0192002 0.0195012 5.4326 44.873 2.630
   10 q     0.100498 0.0187579 0.0189145 5.3133 40.651 2.511
    9 q     0.0966188 0.0164891 0.0166373 5.8073 39.917 2.235
    5 q     0.0870868 0.0167265 0.0168861 5.1573 40.524 2.423
   11 q     0.0859173 0.0161668 0.0162768 5.2785 36.561 2.316
   13 gate  0.106517 0.0188932 0.0190439 5.5932 27.016 1.582
   14 gate  0.106408 0.0191044 0.0192246 5.5350 25.175 1.566
   10 gate  0.102565 0.0183498 0.0183587 5.5867 26.214 1.590
    0 gate  0.0852935 0.0180099 0.0182566 4.6719 24.726 1.638
   13 up    0.103173 0.0186217 0.0187519 5.5020 26.319 1.567
   17 down  0.0884481 0.0160885 0.0161137 5.4890 29.534 1.631
   11 down  0.0858366 0.0175277 0.0176421 4.8655 24.136 1.762
    3 down  0.0653797 0.0166908 0.0167693 3.8988 19.300 1.626
    1 o     0.0795393 0.0166412 0.0166744 4.7701 38.441 2.609
    4 o     0.077728 0.0167011 0.0168708 4.6073 35.113 2.767
   11 k     0.0417367 0.0117871 0.0119861 3.4821 40.120 4.641
   12 v     0.0271273 0.00964749 0.00971123 2.7934 26.642 4.401

=== GATE/UP SUBSPACE CONTROLS ===
layer observed_overlap independently_sign_randomized_overlap
    0 0.971428 0.010807
    4 0.929755 0.016758
   10 0.847464 0.004612
   13 0.972568 0.009036
   17 0.946622 0.013355

=== HEAD-SPECIFIC STRUCTURED DIFFERENCES ===
Only Q_h and O_h here; shared K/V excluded.
s1_rel = leading singular value of difference / RMS paired matrix Frobenius norm.
Ranking is by hypot(Q_s1_rel,O_s1_rel), NOT causal importance.
layer head Q_s1_rel O_s1_rel Q_top1% O_top1% Q_top4% O_top4%
    3    3 0.003818 0.006403 30.294 15.543 54.771 40.574
    4    0 0.003472 0.005450 29.923 17.780 51.241 38.845
    1    2 0.002229 0.005365 17.297 24.102 40.445 48.465
   17    3 0.003082 0.004915 26.202 28.955 47.351 55.984
    4    2 0.002887 0.004799 23.698 16.787 48.963 41.018
    4    1 0.002400 0.005027 15.090 12.652 43.340 32.524
    1    3 0.002816 0.004697 20.800 30.673 57.160 61.150
    4    3 0.003308 0.004136 22.122 24.716 57.292 48.870
    3    1 0.002656 0.004570 16.889 15.312 46.658 39.828
    0    2 0.002599 0.004602 17.565 33.530 48.402 53.631
    2    3 0.003653 0.003759 23.853 11.198 55.509 33.286
   12    0 0.004593 0.002454 41.233 11.463 65.798 36.475
   16    3 0.002561 0.004521 15.714 14.809 37.219 36.795
    7    3 0.002583 0.004498 17.728 17.002 47.923 41.962
    5    2 0.004052 0.003202 36.278 14.469 59.176 37.274
    1    1 0.003049 0.004139 21.635 17.925 54.044 47.594
    8    2 0.002518 0.004451 15.248 14.057 43.175 35.558
    1    0 0.002058 0.004648 15.159 19.751 42.843 53.139
   13    1 0.002889 0.004140 23.434 18.901 46.725 36.525
   13    2 0.003440 0.003607 28.606 14.815 49.958 33.261
    8    1 0.002834 0.004099 19.431 16.358 45.955 35.864
   12    1 0.004130 0.002780 32.102 12.312 56.042 34.963
    5    3 0.003358 0.003575 24.674 15.664 57.467 37.603
    5    1 0.003151 0.003677 20.425 19.134 50.159 42.372

=== MLP UP/DOWN CANCELLING-RESCALING PROJECTION ===
For each unit, use midpoint u,d and the algebraic direction (du,dd)=(a*u,-a*d).
This direction cancels exactly in the midpoint formula for the paired product difference.
Report squared Euclidean projection onto these directions as a fraction of actual
up/down difference energy. This is a closed-form weight-space projection, not training.
layer projected_fraction% total_updown_delta_norm
    0 0.049977 0.378343
    1 0.060392 0.384797
    2 0.064647 0.38306
    3 0.068088 0.374896
    4 0.082121 0.385669
    5 0.053779 0.385717
    6 0.068029 0.379688
    7 0.068686 0.380718
    8 0.070761 0.381191
    9 0.064071 0.384635
   10 0.074228 0.384948
   11 0.062450 0.381622
   12 0.050465 0.395683
   13 0.061463 0.402226
   14 0.057866 0.410287
   15 0.052694 0.411716
   16 0.055943 0.396456
   17 0.069852 0.374603

No network or nonlinear layer was run.

```