OUTPUT of run 24 of at most 60:
```
=== WEIGHT-INTERACTION TENSORS; NO INPUT EVALUATION ===
T2[o,i,j] = sum_u D[o,u] sym(G[u,i] U[u,j])
T3[o,i,j,k] = sum_u D[o,u] sym(G[u,i] G[u,j] U[u,k])
Only weight entries are multiplied. Indices i,j,k are matrix-column
indices, not tokens, hidden states, or synthetic model inputs.
No GELU or other activation function is evaluated.
These tensors are invariant to unit permutation and U-row/D-column
reciprocal rescaling, but do not describe the entire nonlinear network.
Task and vocabulary-control output ranks: 42 42
Importance sampling estimates squared tensor norms from coordinate entries.
Two independent batches assess sampling sensitivity, not behavioural uncertainty.

=== ALL LAYERS ===
Each order: relative tensor change, task/control enrichment,
output-term/total norm, input-term/total norm, batch half-gap.
Term norm ratios need not add to one because terms can reinforce/cancel.
L geo | order2 rel enrich out in gap | order3 rel enrich out in gap
 0 raw  | 0.0133   1.05 0.792 0.604 0.4% | 0.0110   1.09 0.744 0.664 0.6%
 0 diag | 0.0145   0.88 0.842 0.530 0.2% | 0.0112   0.89 0.795 0.613 1.4%
 1 raw  | 0.0202   1.12 0.898 0.432 0.1% | 0.0207   1.09 0.862 0.506 0.1%
 1 diag | 0.0233   1.11 0.927 0.372 0.8% | 0.0237   1.10 0.900 0.437 0.1%
 2 raw  | 0.0180   1.12 0.868 0.490 0.2% | 0.0187   1.11 0.827 0.557 0.9%
 2 diag | 0.0174   1.09 0.863 0.498 0.7% | 0.0184   1.10 0.825 0.563 0.1%
 3 raw  | 0.0183   0.98 0.874 0.477 0.2% | 0.0188   0.99 0.854 0.517 0.8%
 3 diag | 0.0195   0.97 0.856 0.510 0.6% | 0.0190   0.95 0.859 0.510 0.2%
 4 raw  | 0.0203   1.06 0.886 0.454 0.5% | 0.0201   1.05 0.886 0.465 0.5%
 4 diag | 0.0232   0.91 0.924 0.379 0.4% | 0.0244   0.90 0.934 0.356 0.9%
 5 raw  | 0.0143   0.97 0.747 0.652 0.6% | 0.0145   0.96 0.750 0.661 0.5%
 5 diag | 0.0139   0.90 0.738 0.663 0.7% | 0.0139   0.91 0.729 0.680 0.7%
 6 raw  | 0.0177   0.98 0.846 0.523 0.4% | 0.0165   1.04 0.882 0.470 0.2%
 6 diag | 0.0182   0.90 0.852 0.513 0.5% | 0.0196   0.97 0.888 0.457 3.8%
 7 raw  | 0.0168   0.98 0.825 0.558 0.6% | 0.0151   1.02 0.881 0.475 1.3%
 7 diag | 0.0163   0.86 0.826 0.552 0.5% | 0.0113   0.90 0.877 0.481 2.4%
 8 raw  | 0.0207   1.01 0.879 0.464 0.1% | 0.0202   0.98 0.899 0.437 0.3%
 8 diag | 0.0216   0.90 0.890 0.441 1.3% | 0.0215   0.91 0.905 0.423 2.0%
 9 raw  | 0.0165   0.93 0.792 0.598 0.3% | 0.0169   0.90 0.828 0.559 2.4%
 9 diag | 0.0167   0.84 0.801 0.576 1.0% | 0.0169   0.85 0.828 0.559 0.6%
10 raw  | 0.0196   1.04 0.867 0.487 0.2% | 0.0195   1.05 0.889 0.457 5.6%
10 diag | 0.0192   0.89 0.920 0.386 2.1% | 0.0182   1.04 0.963 0.262 8.4%
11 raw  | 0.0164   1.09 0.820 0.563 0.7% | 0.0164   1.10 0.823 0.565 0.2%
11 diag | 0.0161   1.53 0.834 0.531 1.1% | 0.0151   1.47 0.845 0.529 1.5%
12 raw  | 0.0168   1.27 0.802 0.584 0.9% | 0.0177   1.30 0.817 0.574 0.1%
12 diag | 0.0173   1.74 0.825 0.539 1.8% | 0.0159   1.95 0.888 0.476 14.1%
13 raw  | 0.0228   1.43 0.897 0.428 0.0% | 0.0225   1.49 0.900 0.433 0.5%
13 diag | 0.0226   1.86 0.897 0.414 1.6% | 0.0176   2.05 0.910 0.406 3.1%
14 raw  | 0.0304   1.93 0.934 0.338 0.2% | 0.0328   1.99 0.944 0.326 0.5%
14 diag | 0.0297   2.08 0.935 0.330 1.7% | 0.0267   1.93 0.958 0.297 1.4%
15 raw  | 0.0305   2.73 0.929 0.341 0.3% | 0.0301   2.72 0.938 0.343 1.3%
15 diag | 0.0308   2.86 0.929 0.335 0.6% | 0.0305   3.22 0.949 0.314 0.0%
16 raw  | 0.0343   4.10 0.945 0.295 0.0% | 0.0351   4.29 0.959 0.279 5.4%
16 diag | 0.0345   4.34 0.936 0.290 0.1% | 0.0287   4.47 0.958 0.281 1.1%
17 raw  | 0.0422   5.62 0.959 0.246 0.2% | 0.0423   5.80 0.971 0.236 0.8%
17 diag | 0.0476   5.18 0.952 0.235 0.8% | 0.0471   5.32 0.975 0.211 5.1%

=== ROBUST TASK-INTERACTION ENRICHMENT ===
Ranking uses the minimum enrichment across both geometries and orders.
L17: minimum enrichment=5.181, maximum batch half-gap=5.1%
L16: minimum enrichment=4.103, maximum batch half-gap=5.4%
L15: minimum enrichment=2.718, maximum batch half-gap=1.3%
L14: minimum enrichment=1.926, maximum batch half-gap=1.7%
L13: minimum enrichment=1.425, maximum batch half-gap=3.1%
L12: minimum enrichment=1.271, maximum batch half-gap=14.1%
L11: minimum enrichment=1.090, maximum batch half-gap=1.5%
L01: minimum enrichment=1.088, maximum batch half-gap=0.8%
L02: minimum enrichment=1.087, maximum batch half-gap=0.9%
L03: minimum enrichment=0.952, maximum batch half-gap=0.8%
L08: minimum enrichment=0.904, maximum batch half-gap=2.0%
L05: minimum enrichment=0.899, maximum batch half-gap=0.7%
L04: minimum enrichment=0.898, maximum batch half-gap=0.9%
L06: minimum enrichment=0.896, maximum batch half-gap=3.8%
L10: minimum enrichment=0.885, maximum batch half-gap=8.4%
L00: minimum enrichment=0.884, maximum batch half-gap=1.4%
L07: minimum enrichment=0.858, maximum batch half-gap=2.4%
L09: minimum enrichment=0.845, maximum batch half-gap=2.4%

=== DOMAIN CONTENT OF LATE-LAYER INTERACTION CHANGES ===
Per-domain squared tensor norm divided by domain rank, relative to
the control's mean per-output-direction squared tensor norm.
One control subspace and sampled entries make these exploratory estimates.
L11 order2: {'digits': 2.09, 'colors': 1.52, 'regions': 1.43, 'cities': 1.42, 'languages': 1.51}
L11 order3: {'digits': 1.9, 'colors': 1.45, 'regions': 1.5, 'cities': 1.34, 'languages': 1.45}
L12 order2: {'digits': 2.78, 'colors': 1.66, 'regions': 1.62, 'cities': 1.53, 'languages': 1.47}
L12 order3: {'digits': 2.57, 'colors': 1.69, 'regions': 3.33, 'cities': 1.69, 'languages': 1.42}
L13 order2: {'digits': 3.61, 'colors': 1.71, 'regions': 1.76, 'cities': 1.37, 'languages': 1.37}
L13 order3: {'digits': 3.19, 'colors': 1.94, 'regions': 2.72, 'cities': 1.59, 'languages': 1.95}
L14 order2: {'digits': 5.17, 'colors': 1.84, 'regions': 1.69, 'cities': 1.12, 'languages': 1.2}
L14 order3: {'digits': 4.72, 'colors': 1.45, 'regions': 1.47, 'cities': 0.96, 'languages': 1.41}
L15 order2: {'digits': 8.45, 'colors': 2.05, 'regions': 2.32, 'cities': 1.24, 'languages': 1.24}
L15 order3: {'digits': 7.66, 'colors': 3.41, 'regions': 4.61, 'cities': 1.31, 'languages': 1.54}
L16 order2: {'digits': 14.01, 'colors': 3.06, 'regions': 3.24, 'cities': 1.56, 'languages': 1.5}
L16 order3: {'digits': 12.93, 'colors': 4.58, 'regions': 4.21, 'cities': 1.69, 'languages': 1.99}
L17 order2: {'digits': 16.08, 'colors': 4.42, 'regions': 5.05, 'cities': 2.58, 'languages': 2.14}
L17 order3: {'digits': 15.36, 'colors': 4.74, 'regions': 6.6, 'cities': 2.96, 'languages': 2.4}

```