OUTPUT of run 3 of at most 60:
```
Attention configuration: {'num_attention_heads': 4, 'num_key_value_heads': 1, 'head_dim': 256, 'hidden_size': 640, 'attention_bias': False}
Using standard contiguous grouped-query head assignment.

PER-LAYER RESULTS:
layer=0 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0103308
layer=1 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0138906
layer=2 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0163669
layer=3 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0218453
layer=4 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0208619
layer=5 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0145085
layer=6 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0206074
layer=7 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0241073
layer=8 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0237504
layer=9 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0264315
layer=10 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0129635
layer=11 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.026559
layer=12 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.014559
layer=13 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0182503
layer=14 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0136138
layer=15 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0254158
layer=16 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0258957
layer=17 changed_corresponding_heads=4/4 A_heads_matching_no_B_head=4/4 nearest_B_heads=[0, 1, 2, 3] summed_map_relative_delta=0.0153287

TOP 20 HEADS BY RELATIVE MAP CHANGE — NOT CAUSAL IMPORTANCE:
L3 H3 KV0 relative_delta=0.0188356 extraction/delta=0.3629 writing/delta=0.8998 term_cosine=0.0898
L4 H1 KV0 relative_delta=0.0176839 extraction/delta=0.3854 writing/delta=0.8630 term_cosine=0.1604
L4 H0 KV0 relative_delta=0.016224 extraction/delta=0.4138 writing/delta=0.8726 term_cosine=0.0932
L16 H3 KV0 relative_delta=0.015897 extraction/delta=0.5340 writing/delta=0.8030 term_cosine=0.0817
L16 H1 KV0 relative_delta=0.0155487 extraction/delta=0.5859 writing/delta=0.7635 term_cosine=0.0824
L4 H2 KV0 relative_delta=0.014252 extraction/delta=0.4379 writing/delta=0.8546 term_cosine=0.1042
L8 H2 KV0 relative_delta=0.0141482 extraction/delta=0.4651 writing/delta=0.8518 term_cosine=0.0732
L3 H1 KV0 relative_delta=0.0140597 extraction/delta=0.4558 writing/delta=0.8410 term_cosine=0.1109
L16 H0 KV0 relative_delta=0.0139915 extraction/delta=0.6255 writing/delta=0.7471 term_cosine=0.0542
L7 H3 KV0 relative_delta=0.0136503 extraction/delta=0.4967 writing/delta=0.8303 term_cosine=0.0774
L5 H3 KV0 relative_delta=0.0132809 extraction/delta=0.5904 writing/delta=0.7557 term_cosine=0.0900
L2 H3 KV0 relative_delta=0.0129609 extraction/delta=0.5031 writing/delta=0.8147 term_cosine=0.1014
L1 H0 KV0 relative_delta=0.0129155 extraction/delta=0.4759 writing/delta=0.8289 term_cosine=0.1097
L15 H2 KV0 relative_delta=0.0129154 extraction/delta=0.6523 writing/delta=0.6986 term_cosine=0.0950
L11 H1 KV0 relative_delta=0.0127736 extraction/delta=0.5384 writing/delta=0.7680 term_cosine=0.1454
L13 H1 KV0 relative_delta=0.012744 extraction/delta=0.5887 writing/delta=0.7484 term_cosine=0.1058
L3 H2 KV0 relative_delta=0.0126432 extraction/delta=0.4698 writing/delta=0.8180 term_cosine=0.1433
L5 H2 KV0 relative_delta=0.01261 extraction/delta=0.5806 writing/delta=0.7709 term_cosine=0.0766
L8 H1 KV0 relative_delta=0.0126031 extraction/delta=0.5401 writing/delta=0.8105 term_cosine=0.0587
L12 H3 KV0 relative_delta=0.0125731 extraction/delta=0.7010 writing/delta=0.6639 term_cosine=0.0729

SUMMARY:
Heads tested: 72
Heads with resolved product differences: 72
Layers ruling out preservation by any head permutation: [0, 1, 10, 11, 12, 13, 14, 15, 16, 17, 2, 3, 4, 5, 6, 7, 8, 9]
Smallest difference / roundoff estimate: 257378226.32799643
Largest relative decomposition error: 8.635388785997712e-18
Skipped: []

Scope: these are linear value/output products, not evaluated attention functions. Their differences do not establish routing, whole-block differences after compensation, semantic content, behavioral importance, or the sycophantic checkpoint.

```