OUTPUT of run 4 of at most 60:
```
CROSS-MODULE OUTPUT-GEOMETRY TEST
Blocks: 36 residual coordinates: 640
Matrices: attention O projections and MLP down projections.
These are writer matrices, not complete attention or MLP functions.
No post-block state-dependent normalization is included.
Coherence of 1 is the sign-randomization expectation.
Larger values indicate coordinated off-diagonal covariance geometry.

DELTA GEOMETRY VERSUS MIDPOINT GEOMETRY
Per-block cosine min/median/max: [-0.0576 -0.022   0.0113]
Residual geometry norm after removing midpoint-aligned component:
min/median/max: [0.9983 0.9998 1.    ]
Inactive output rows in delta: 0

TEST 1: alignment of checkpoint-difference output geometry
Observed coherence: 3.35865597853324
Mean pairwise Frobenius cosine: 0.06739017081523542
Sign-permuted coherence min/median/max: [0.9922 0.9993 1.0065]
Sign-permuted coherence mean/std: 0.9993907448103598 0.002839891335981634
Controls >= observed: 0 out of 64
This is a structural surrogate comparison, not a behavioural p-value.
Largest contributors to cross-block alignment:
  L15 mlp       alignment_to_others=3.748343 delta_midpoint_cosine=-0.01480996
  L14 mlp       alignment_to_others=3.494895 delta_midpoint_cosine=-0.03737227
  L16 mlp       alignment_to_others=3.456414 delta_midpoint_cosine=-0.008897405
  L16 attention alignment_to_others=3.209442 delta_midpoint_cosine=-0.03319185
  L08 mlp       alignment_to_others=3.093451 delta_midpoint_cosine=-0.02041098
  L15 attention alignment_to_others=3.052984 delta_midpoint_cosine=-0.02457825
  L07 mlp       alignment_to_others=2.955531 delta_midpoint_cosine=-0.01225704
  L13 mlp       alignment_to_others=2.925845 delta_midpoint_cosine=-0.03472367
  L09 mlp       alignment_to_others=2.828717 delta_midpoint_cosine=-0.023566
  L06 mlp       alignment_to_others=2.824337 delta_midpoint_cosine=-0.007633825

TEST 2: alignment after removing midpoint-parallel geometry
Observed coherence: 3.354462333987401
Mean pairwise Frobenius cosine: 0.06727035239964003
Sign-permuted coherence min/median/max: [0.9921 0.9993 1.0064]
Sign-permuted coherence mean/std: 0.9993888075069056 0.002837968451045027
Controls >= observed: 0 out of 64
This is a structural surrogate comparison, not a behavioural p-value.
Largest contributors to cross-block alignment:
  L15 mlp       alignment_to_others=3.74268 delta_midpoint_cosine=-0.01480996
  L14 mlp       alignment_to_others=3.483417 delta_midpoint_cosine=-0.03737227
  L16 mlp       alignment_to_others=3.452467 delta_midpoint_cosine=-0.008897405
  L16 attention alignment_to_others=3.200138 delta_midpoint_cosine=-0.03319185
  L08 mlp       alignment_to_others=3.089223 delta_midpoint_cosine=-0.02041098
  L15 attention alignment_to_others=3.047134 delta_midpoint_cosine=-0.02457825
  L07 mlp       alignment_to_others=2.954479 delta_midpoint_cosine=-0.01225704
  L13 mlp       alignment_to_others=2.917025 delta_midpoint_cosine=-0.03472367
  L09 mlp       alignment_to_others=2.826464 delta_midpoint_cosine=-0.023566
  L06 mlp       alignment_to_others=2.822889 delta_midpoint_cosine=-0.007633825

SHARED ROW-STANDARDIZED OUTPUT COVARIANCE
Top 8 delta eigenvalues: [10.329   8.648   7.2431  6.9134  6.6645  6.4564  5.7458  5.6549]
Top 8 midpoint eigenvalues: [1.8538 1.795  1.7444 1.7013 1.676  1.6685 1.6396 1.6342]
Top-8 delta trace fraction: 0.09008602659214926
Top-8 midpoint trace fraction: 0.02142611143638748
Top-8 subspace overlap, 0 to 1: 0.019834382139960468
Random independent subspace reference k/d: 0.0125
Largest coordinates in leading delta eigenvector, with squared loading:
[(263, 0.02492002299383768), (4, 0.02383775427818844), (153, 0.020045622304131153), (328, 0.014573327274604135), (300, 0.012902989782722469), (486, 0.012641928699059723), (71, 0.012616029050970784), (29, 0.010580050210551759), (399, 0.010448143567945437), (576, 0.010434537431289503), (350, 0.00978879194872495), (193, 0.009009656051503858)]

Interpretation boundary:
Positive alignment would identify a coordinated weight-change geometry.
It would not identify whether that geometry encodes agreement, correctness,
formatting, or another training-related change.
The test is unchanged by exchanging A and B.

```