OUTPUT of run 7 of at most 60:
```
=== NEARBY WEIGHT-PATH SUBSPACE CONNECTIONS ===
Recipients: joint gate/up reading subspaces, rank 8.
Attention sources: left subspaces of paired OV differences.
MLP sources: left subspaces of down-projection differences.
Only direct residual-coordinate compatibility is tested.
Unknown gates, attention weights, normalization denominators and intervening
branch contributions are NOT supplied or simulated.
The sign-null controls coordinate concentration, not training statistics.

Top 20 connections by MIN(raw excess overlap, scaled excess overlap).
This conservative structural ranking is not a causal circuit ranking.
source -> target raw_overlap raw_null scaled_overlap scaled_null scaled_baseread_overlap scaled_baseread_null
L08H2   -> L08MLP 0.03435 0.01238 0.01094 0.00328 0.00209 0.00102
L08H3   -> L08MLP 0.02887 0.01225 0.00944 0.00344 0.00226 0.00141
L07H3   -> L08MLP 0.02656 0.01249 0.00806 0.00293 0.00096 0.00073
L08H1   -> L08MLP 0.02853 0.01224 0.00839 0.00340 0.00267 0.00140
L08H2   -> L09MLP 0.04877 0.01264 0.00648 0.00150 0.00026 0.00015
L09H2   -> L09MLP 0.04412 0.01255 0.00617 0.00168 0.00031 0.00016
L06H0   -> L06MLP 0.02649 0.01265 0.00736 0.00300 0.00432 0.00166
L08H1   -> L09MLP 0.04629 0.01250 0.00561 0.00152 0.00024 0.00014
L07MLP  -> L08MLP 0.02928 0.01264 0.00675 0.00294 0.00138 0.00155
L08MLP  -> L09MLP 0.05083 0.01246 0.00503 0.00129 0.00025 0.00013
L07H2   -> L08MLP 0.02736 0.01249 0.00617 0.00267 0.00097 0.00066
L07H3   -> L09MLP 0.02978 0.01241 0.00522 0.00173 0.00054 0.00030
L09H3   -> L09MLP 0.03999 0.01217 0.00498 0.00156 0.00037 0.00015
L05MLP  -> L06MLP 0.03150 0.01240 0.00489 0.00152 0.00189 0.00070
L06MLP  -> L08MLP 0.03144 0.01263 0.00561 0.00231 0.00094 0.00060
L04H2   -> L05MLP 0.02883 0.01244 0.00521 0.00193 0.00156 0.00108
L08H3   -> L09MLP 0.04131 0.01238 0.00471 0.00151 0.00021 0.00014
L07H1   -> L08MLP 0.02482 0.01240 0.00598 0.00293 0.00069 0.00078
L08H0   -> L08MLP 0.02182 0.01197 0.00697 0.00391 0.00565 0.00354
L06H2   -> L06MLP 0.03149 0.01244 0.00594 0.00292 0.00313 0.00151
Overlap of raw/scaled top-20 connections: 3

=== SIGNED IDENTITY ALIGNMENT OF OV ===
c(M)=trace(M)/(sqrt(hidden_size)*||M||_F).
A purely algebraic coordinate-preservation measure, NOT user-guess copying.
scaled= False B>A: 34 B<A: 38 change[min,median,max]= [-0. -0.  0.]
scaled= True B>A: 36 B<A: 36 change[min,median,max]= [-0.0002  0.      0.0001]

Selected diagonally scaled operators:
layer head c_A c_B B_minus_A min_eig(sym_delta) max_eig(sym_delta)
    3    3 -0.00846830 -0.00847668 -0.00000837 -16.6928 16.6655
    4    0 -0.00012984 -0.00012306 +0.00000677 -31.2175 31.6446
    4    1 0.00663938 0.00666121 +0.00002183 -35.0019 34.8745
    8    2 0.13464572 0.13466817 +0.00002245 -20.1486 20.5276
   10    1 0.01138335 0.01138118 -0.00000217 -39.3982 39.6519
   11    0 0.00253390 0.00253280 -0.00000111 -52.1125 52.7644
   11    1 0.05363285 0.05363244 -0.00000041 -36.2856 36.2336
   12    0 -0.00684178 -0.00684006 +0.00000173 -177.472 177.793
   12    2 0.00982985 0.00982927 -0.00000058 -196.103 197.011
   17    3 -0.00602858 -0.00603090 -0.00000232 -107.313 105.842

Opposite eigenvalue signs mean direction-dependent symmetric change,
not that either model uniformly strengthens coordinate-preserving transport.
No residual direction has been identified with guesses or correct answers.

```