OUTPUT of run 7 of at most 60:
```
=== DIRECT WEIGHT-PATH CONNECTIVITY TEST ===
Selectivity = alignment with changed-output covariance / alignment with baseline-output covariance.
Both covariances have trace 1. Percentile compares all 2048 units in the same target layer/branch.
This is a descriptive comparison, not a statistical significance test.
source -> target unit | gate selectivity percentile | up selectivity percentile | prior shortlist
 1 ->  2 1831 |    2.3526   96.88 |    1.8704   87.79 | True
 2 ->  2 1831 |    1.1066   96.14 |    0.8845   86.82 | True
 2 ->  3 1414 |    0.6999   54.93 |    0.5417   17.82 | True
 3 ->  3 1414 |    0.7720   41.80 |    0.7611   47.66 | True
 2 ->  3  226 |    0.6135   30.08 |    0.8922   93.55 | True
 3 ->  3  226 |    0.7839   43.55 |    1.2945   91.85 | True
 2 ->  3  986 |    0.6475   39.94 |    0.7698   79.64 | True
 3 ->  3  986 |    0.7982   45.65 |    0.9737   76.03 | True
 2 ->  3   82 |    0.5919   24.27 |    0.6999   65.67 | True
 3 ->  3   82 |    0.4149    2.20 |    0.5224    8.74 | True
 9 -> 10 1363 |    1.0894   72.75 |    0.8412   44.48 | True
10 -> 10 1363 |    3.9847   91.55 |    2.4596   77.29 | True
11 -> 12 1528 |    0.1893    0.10 |    0.4375    1.03 | False
12 -> 12 1528 |    0.6070    5.66 |    0.7838   18.12 | False
11 -> 12 1520 |    1.3136   61.96 |    1.9451   94.78 | False
12 -> 12 1520 |    2.5721   91.80 |    2.1601   92.97 | False
11 -> 12  259 |    2.9281   96.58 |    1.5350   84.86 | False
12 -> 12  259 |    2.2124   87.26 |    2.2813   93.99 | False
13 -> 14  641 |    3.2072   97.66 |    2.1589   87.70 | True
14 -> 14  641 |    0.3232    0.10 |    2.2392   78.91 | True
14 -> 15 1369 |    2.8075   95.85 |    1.3407   21.73 | False
15 -> 15 1369 |    0.1180    0.05 |    0.2533    4.83 | False
15 -> 16  893 |    0.6406   84.52 |    1.1084   95.65 | False
16 -> 16  893 |    2.8542   95.95 |    4.7060   97.51 | False

=== CONNECTIVITY SUMMARY FOR PREVIOUSLY SHORTLISTED UNITS ===
Number of unit/source/branch comparisons: 28
Median percentile: 69.2138671875
Fraction above 90th percentile: 0.21428571428571427
Fraction below 10th percentile: 0.10714285714285714
Comparisons are correlated and were not selected independently of the weights.

=== CROSS-COMPONENT REINFORCEMENT/CANCELLATION ===
Ratio = ||attention_term + reader_term||^2 / (||attention_term||^2 + ||reader_term||^2).
1 means zero cross term; below 1 means cancellation; above 1 means reinforcement.
Number of selected paths: 192
Ratio min / 10% / median / 90% / max: [0.5961 0.8507 1.0099 1.2006 1.4194]
Fraction below 0.5: 0.0
Fraction above 1.5: 0.0
Maximum exact-decomposition reconstruction error: 3.102897414786613e-13

=== STRONGEST CANCELLATION AMONG SELECTED PATHS ===
attn_layer head -> MLP_layer unit branch | relative_path_change reinforcement | attention_share reader_share
10 0 -> 10 1363 gate | 0.022715 0.596066 | +0.360446 +0.639554
12 1 -> 12 1520 up   | 0.021966 0.654129 | +0.851112 +0.148888
14 3 -> 15 1369 up   | 0.009892 0.674833 | +0.696573 +0.303427
11 2 -> 12 1520 up   | 0.017235 0.697641 | +0.454339 +0.545661
11 3 -> 12  259 gate | 0.017544 0.697923 | +0.834563 +0.165437
12 2 -> 12 1520 up   | 0.024917 0.759502 | +0.181050 +0.818950
10 0 -> 10 1363 up   | 0.023118 0.798302 | +0.242663 +0.757337
14 0 -> 14  641 up   | 0.012485 0.804237 | +0.614508 +0.385492
11 1 -> 12 1520 up   | 0.014540 0.810565 | +0.558587 +0.441413
 1 1 ->  2 1831 gate | 0.016367 0.816223 | +1.002149 -0.002149

=== STRONGEST REINFORCEMENT AMONG SELECTED PATHS ===
attn_layer head -> MLP_layer unit branch | relative_path_change reinforcement | attention_share reader_share
12 2 -> 12  259 up   | 0.039972 1.419417 | +0.555328 +0.444672
11 3 -> 12 1520 up   | 0.023611 1.381752 | +0.428704 +0.571296
12 3 -> 12  259 up   | 0.034740 1.321161 | +0.421624 +0.578376
13 1 -> 14  641 gate | 0.025639 1.312741 | +0.774107 +0.225893
11 2 -> 12  259 up   | 0.022094 1.303685 | +0.512393 +0.487607
12 3 -> 12  259 gate | 0.033248 1.297692 | +0.651497 +0.348503
12 1 -> 12  259 gate | 0.029408 1.289637 | +0.775962 +0.224038
 3 0 ->  3   82 up   | 0.029513 1.288576 | +0.778916 +0.221084
13 1 -> 14  641 up   | 0.033119 1.272184 | +0.645895 +0.354105
15 0 -> 15 1369 up   | 0.011551 1.254389 | +0.582681 +0.417319

=== LARGEST RELATIVE COMPOSED-PATH CHANGES ===
attn_layer head -> MLP_layer unit branch | relative_path_change reinforcement | attention_share reader_share
 3 1 ->  3   82 gate | 0.065080 1.132652 | +0.917575 +0.082425
 3 0 ->  3   82 gate | 0.053741 1.141225 | +0.899834 +0.100166
 3 2 ->  3   82 gate | 0.053144 1.122675 | +0.913513 +0.086487
 3 1 ->  3   82 up   | 0.040179 1.206315 | +0.785781 +0.214219
12 2 -> 12  259 up   | 0.039972 1.419417 | +0.555328 +0.444672
16 1 -> 16  893 up   | 0.038439 1.022823 | +0.945550 +0.054450
12 1 -> 12  259 up   | 0.036406 1.248684 | +0.688275 +0.311725
12 2 -> 12  259 gate | 0.034864 1.209670 | +0.711303 +0.288697
12 3 -> 12  259 up   | 0.034740 1.321161 | +0.421624 +0.578376
12 3 -> 12  259 gate | 0.033248 1.297692 | +0.651497 +0.348503
13 1 -> 14  641 up   | 0.033119 1.272184 | +0.645895 +0.354105
 2 3 ->  3   82 gate | 0.032935 1.039582 | +0.947670 +0.052330
 1 2 ->  2 1831 gate | 0.030831 0.846153 | +1.016410 -0.016410
16 3 -> 16  893 gate | 0.029970 0.988168 | +0.978881 +0.021119
 3 3 ->  3  226 up   | 0.029782 1.020258 | +0.936940 +0.063060
 3 2 ->  3  226 up   | 0.029574 1.185804 | +0.787302 +0.212698

Interpretation limits:
- These are potential linear connections, not executed attention-to-MLP pathways.
- Same-layer attention precedes its MLP; next-layer comparisons follow a residual-skip connection.
- Intervening computations, normalization denominators, attention probabilities and GELU gates are omitted.
- Reinforcement of weight products does not imply reinforcement of a behavioral effect.
- No inputs, activations, logits, or nonlinear model computations were used.

```