OUTPUT of run 9 of at most 100:
```
L0 DOMINANT QK FACTORS
H0: top-mode energy=0.8877, query-pos32=0.8954, key-pos32=0.2819
H1: top-mode energy=0.7148, query-pos32=0.7432, key-pos32=0.4198
H2: top-mode energy=0.1204, query-pos32=0.9868, key-pos32=0.9970
H3: top-mode energy=0.8269, query-pos32=0.8838, key-pos32=0.9328
H4: top-mode energy=0.7035, query-pos32=0.5431, key-pos32=0.4386
H5: top-mode energy=0.9578, query-pos32=0.8676, key-pos32=0.5002
absolute cosines of leading query directions
[[1.     0.77   0.4272 0.7972 0.1273 0.2389]
 [0.77   1.     0.5194 0.7382 0.227  0.0475]
 [0.4272 0.5194 1.     0.4817 0.1129 0.0606]
 [0.7972 0.7382 0.4817 1.     0.0648 0.0329]
 [0.1273 0.227  0.1129 0.0648 1.     0.3261]
 [0.2389 0.0475 0.0606 0.0329 0.3261 1.    ]]
absolute cosines of leading key directions
[[1.     0.264  0.0526 0.0032 0.3793 0.0979]
 [0.264  1.     0.0191 0.1238 0.2577 0.1429]
 [0.0526 0.0191 1.     0.513  0.0234 0.1277]
 [0.0032 0.1238 0.513  1.     0.101  0.0119]
 [0.3793 0.2577 0.0234 0.101  1.     0.6433]
 [0.0979 0.1429 0.1277 0.0119 0.6433 1.    ]]

POSITION-SUBSPACE AND AFFINE WEIGHT COEFFICIENTS
query-pos32 fraction; key-linear norm / QK norm; key-linear fraction in common 3D writes
L1H0 [0.105  1.5297 0.8698]
L1H1 [0.1152 1.7366 0.9031]
L1H2 [0.1234 1.6683 0.9041]
L1H3 [0.1198 1.7087 0.8899]
L1H4 [0.1231 1.5788 0.9054]
L1H5 [0.1265 1.3096 0.8894]
L2H0 [0.0826 0.6969 0.8566]
L2H1 [0.0713 0.3601 0.7939]
L2H2 [0.0855 0.4684 0.8053]
L2H3 [0.08   0.7009 0.8732]
L2H4 [0.0852 0.6729 0.8644]
L2H5 [0.0807 0.6703 0.8574]
L3H0 [0.0772 0.1156 0.1503]
L3H1 [0.0673 0.1319 0.0366]
L3H2 [0.0797 0.1923 0.5941]
L3H3 [0.0831 0.2253 0.6708]
L3H4 [0.0801 0.1655 0.52  ]
L3H5 [0.0777 0.1971 0.0378]

ACTIVE L0 MLP WEIGHT-ROW CLUSTERS
cosine>0.8: components=284, non-singleton neurons=183, largest sizes=[7 5 5 5 5 5 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 3 3 3 3 3 3 3]
cosine>0.9: components=331, non-singleton neurons=127, largest sizes=[6 4 4 4 4 4 4 4 4 4 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 2 2 2 2 2]
cosine>0.95: components=355, non-singleton neurons=95, largest sizes=[4 4 4 4 4 4 3 3 3 3 3 3 3 3 3 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2]
cosine>0.98: components=409, non-singleton neurons=8, largest sizes=[2 2 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
Previous-write top24 read-energy fraction quantiles [0.0058 0.0196 0.0675 0.1157 0.1704 0.2296 0.7185]
Position top32 read-energy fraction quantiles [0.0021 0.0058 0.0167 0.0289 0.0376 0.0515 0.0909]

```