OUTPUT of run 6 of at most 100:
```
Final-layer projection spectrum
[28.461 22.214 21.34  20.708 19.943 19.662 18.986 18.744 18.295 17.457 17.132 16.561 15.329 14.874 14.328 13.438 12.667 12.455 12.317 11.531 11.047
  9.084  8.353  8.24   6.37   5.836  4.626  4.583  3.777  3.61   3.252  3.183  3.086  2.941  2.865  2.801  2.639  2.564  2.538  2.466]
Leading output directions: participation ratio, max absolute coordinate, mean-direction cosine
0 [12.9537  0.428  -0.0207] [(184, -0.428), (39, -0.342), (5, 0.321), (383, -0.254), (132, -0.231), (219, -0.217), (21, -0.216), (89, -0.215)]
1 [117.4406   0.1726   0.0018] [(351, -0.173), (316, 0.161), (283, -0.151), (43, 0.139), (194, -0.138), (366, 0.136), (147, 0.134), (247, 0.126)]
2 [137.8838   0.1599  -0.0017] [(323, 0.16), (219, -0.126), (95, 0.123), (61, -0.123), (201, -0.122), (184, -0.117), (49, 0.116), (367, -0.115)]
3 [124.7351   0.1654  -0.0019] [(90, -0.165), (200, -0.16), (8, -0.149), (355, 0.134), (383, -0.131), (53, 0.129), (174, -0.127), (128, 0.121)]
4 [140.5561   0.1374  -0.0141] [(161, 0.137), (135, 0.131), (10, 0.131), (302, -0.124), (182, -0.124), (284, -0.123), (252, 0.122), (114, -0.121)]
5 [142.0652   0.1433   0.0124] [(36, 0.143), (300, 0.143), (26, -0.142), (124, 0.128), (234, -0.122), (110, 0.12), (269, 0.118), (199, -0.115)]

Final-layer output-column cosine quantiles with leading output direction
0 [-0.922 -0.815 -0.525 -0.312 -0.067  0.135  0.319  0.565  0.753]
1 [-0.792 -0.588 -0.359 -0.181 -0.001  0.175  0.338  0.534  0.774]
2 [-0.669 -0.571 -0.333 -0.188 -0.013  0.171  0.325  0.534  0.683]

First-layer MLP coefficient clustering in H3's leading input-weight subspace
k 8 silhouette 0.0444 sizes [343 292 238 158 152 151 123  79]
k 12 silhouette 0.0565 sizes [288 209 160 138 124 108  98  97  89  89  71  65]
k 16 silhouette 0.068 sizes [274 186 124 105  99  84  83  82  78  76  61  61  60  59  59  45]
k 20 silhouette 0.0727 sizes [166 142 129 127  94  78  74  70  67  64  63  60  60  56  55  51  48  47  44  41]
k 24 silhouette 0.0837 sizes [119 111 111 108 106  99  91  63  58  57  56  55  55  55  54  52  48  40  37  36  36  32  29  28]
k 28 silhouette 0.083 sizes [124 115 105  99  85  74  61  61  53  52  52  51  46  45  44  43  43  41  40  39  39  38  36  34  33  28  28  27]
k 32 silhouette 0.0883 sizes [93 93 89 84 70 66 65 64 57 52 49 48 48 43 43 43 43 41 40 39 39 39 37 35 32 31 30 27 27 27 22 20]

First-layer MLP coefficient row-norm quantiles in current/prior subspaces
16 current [0.044 0.08  0.16  1.216 2.123]
16 prior [0.03  0.047 0.088 1.004 2.316]
20 current [0.05  0.092 0.174 1.321 2.274]
20 prior [0.032 0.05  0.092 1.022 2.331]
24 current [0.056 0.096 0.179 1.337 2.308]
24 prior [0.034 0.051 0.093 1.022 2.332]

L0 output-row norms for low-rank heads and MLP0
H 1 [(39, 2.401), (21, 1.884), (58, 1.773), (99, 1.239), (184, 0.788), (163, 0.328), (153, 0.304), (336, 0.263), (290, 0.258), (377, 0.247)]
H 2 [(39, 1.984), (21, 1.667), (58, 1.495), (99, 1.285), (184, 0.745), (163, 0.261), (217, 0.257), (165, 0.248), (52, 0.246), (98, 0.244)]
H 4 [(39, 1.613), (21, 1.244), (58, 1.135), (99, 0.784), (184, 0.555), (165, 0.24), (52, 0.226), (109, 0.223), (163, 0.22), (153, 0.213)]
H 5 [(21, 5.347), (39, 3.769), (99, 3.356), (58, 3.004), (184, 1.533), (163, 0.493), (354, 0.481), (93, 0.475), (217, 0.462), (81, 0.458)]
MLP0 [(184, 8.562), (21, 5.174), (266, 3.414), (383, 3.25), (294, 3.099), (37, 2.864), (39, 2.858), (60, 2.85), (86, 2.842), (133, 2.8)]

```