OUTPUT of run 8 of at most 100:
```
Final MLP projection-column clustering (weight vectors only)
r 20 k 20 silhouette 0.0655 sizes [106  78  70  63  61  60  60  60  58  57  57  55  55  51  50  50  48  42  38  33]
r 20 k 24 silhouette 0.0637 sizes [96 63 57 57 55 54 52 52 51 50 49 49 47 45 45 43 41 39 38 38 38 37 32 24]
r 20 k 28 silhouette 0.0656 sizes [81 53 50 48 47 47 47 45 45 44 44 41 40 40 40 39 37 37 36 36 36 35 34 33 32 31 31 23]
r 24 k 20 silhouette 0.0558 sizes [87 76 71 67 66 66 65 65 57 56 53 52 52 51 51 50 49 46 37 35]
r 24 k 24 silhouette 0.0595 sizes [103  63  62  59  59  59  54  51  50  49  49  48  47  42  42  41  40  39  39  37  35  29  29  26]
r 24 k 28 silhouette 0.0575 sizes [74 62 53 52 51 51 50 48 45 45 45 45 44 43 41 39 39 37 36 35 35 30 28 28 26 26 23 21]
r 28 k 20 silhouette 0.0583 sizes [84 67 65 65 63 62 61 60 59 59 55 55 55 55 53 50 50 47 45 42]
r 28 k 24 silhouette 0.0604 sizes [104  65  57  57  55  53  53  50  47  47  47  47  47  46  45  43  41  40  39  38  36  32  32  31]
r 28 k 28 silhouette 0.0609 sizes [89 55 51 47 47 47 47 47 46 44 44 43 42 42 41 41 37 37 36 36 36 34 32 29 27 27 25 23]

L3H3 induction composition versus other heads, with rotated-weight null controls
L3H0 trace/norm -0.597 symmetric positive/negative eigenvalue sums [ 0.1106 -0.174 ] null mean/sd [0.009 0.214] null max 0.765
L3H1 trace/norm -0.09 symmetric positive/negative eigenvalue sums [ 0.0713 -0.076 ] null mean/sd [-0.015  0.193] null max 0.607
L3H2 trace/norm 0.184 symmetric positive/negative eigenvalue sums [ 0.0927 -0.0821] null mean/sd [-0.004  0.205] null max 0.514
L3H3 trace/norm 2.73 symmetric positive/negative eigenvalue sums [ 0.4513 -0.0362] null mean/sd [-0.029  0.23 ] null max 0.622
L3H4 trace/norm 0.551 symmetric positive/negative eigenvalue sums [ 0.0426 -0.0287] null mean/sd [-0.007  0.199] null max 0.504
L3H5 trace/norm -0.152 symmetric positive/negative eigenvalue sums [ 0.0969 -0.1077] null mean/sd [0.011 0.202] null max 0.639

```