OUTPUT of run 16 of at most 60:
```
=== MLP UNIT CORRESPONDENCE AND GATE REDUNDANCY ===
Gates include the learned input diagonal scale, but no hidden-state normalization.
Distances compare weight rows and include both orientation and magnitude.
 L A_to_B_same B_to_A_same duplicate_groups_A/B inactive_A/B competitor_margin[min,p01,median]
 0        2048        2048    0/0       0/0    [64.656,103.491,196.980]
 1        2048        2048    0/0       0/0    [84.398,115.800,206.497]
 2        2048        2048    0/0       0/0    [66.713,94.309,203.980]
 3        2048        2048    0/0       0/0    [28.774,40.858,156.804]
 4        2048        2048    0/0       0/0    [37.519,58.067,160.252]
 5        2048        2048    0/0       0/0    [29.780,43.307,159.159]
 6        2048        2048    0/0       0/0    [51.703,73.071,159.798]
 7        2048        2048    0/0       0/0    [61.378,82.659,159.982]
 8        2048        2048    0/0       0/0    [43.831,74.170,149.018]
 9        2048        2048    0/0       0/0    [34.038,54.312,149.257]
10        2048        2048    0/0       0/0    [32.891,58.519,142.304]
11        2048        2048    0/0       0/0    [36.902,55.094,141.121]
12        2048        2048    0/0       0/0    [31.089,45.931,140.953]
13        2048        2048    0/0       0/0    [53.786,86.614,149.304]
14        2048        2048    0/0       0/0    [62.723,85.256,141.862]
15        2048        2048    0/0       0/0    [69.717,91.068,141.740]
16        2048        2048    0/0       0/0    [72.197,92.010,143.505]
17        2048        2048    0/0       0/0    [39.991,76.213,130.264]

=== HIGHLIGHTED UNIT CORRESPONDENCE ===
A competitor margin >1 means the same-index B gate is closer than every other B gate.
 L unit best_B same_index_distance nearest_other_B other_distance margin
 1 1725   1725           0.0030502            2045      1.3374654    438.479
 1 2035   2035           0.0050329             617      1.1843939    235.330
 4   13     13           0.0042147            1757      0.4101840     97.322
 4  937    937           0.0025014             285      1.0803671    431.913
 4 1804   1804           0.0019696            1106      1.0649184    540.665
 4 1044   1044           0.0026837             316      1.1352197    423.002
 5  584    584           0.0060489             200      1.0396213    171.869
 6 1824   1824           0.0004712             560      1.3020394   2763.325
10 1363   1363           0.0009069            1106      1.1318323   1248.039
10 1564   1564           0.0013566            1259      1.1284079    831.808
10 1529   1529           0.0011322            1012      1.2058138   1064.983
13   16     16           0.0017099             829      1.0065116    588.648
14  641    641           0.0010451            1798      1.0542868   1008.832
16  893    893           0.0008981             324      1.0495652   1168.606

=== WITHIN-CHECKPOINT ALTERNATIVES FOR HIGHLIGHTED UNITS ===
Neighbor cosines describe weight geometry, not activation similarity.
L01 unit=1725: A_neighbor=2045, A_distance=1.337016; B_neighbor=2045, B_distance=1.337467; neighbor_gate_cos=+0.123038, neighbor_up_cos=+0.010482; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L01 unit=2035: A_neighbor= 617, A_distance=1.184388; B_neighbor= 617, B_distance=1.184634; neighbor_gate_cos=+0.309173, neighbor_up_cos=+0.230509; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L04 unit=  13: A_neighbor=1757, A_distance=0.410113; B_neighbor=1757, B_distance=0.410194; neighbor_gate_cos=+0.925999, neighbor_up_cos=+0.101003; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L04 unit= 937: A_neighbor= 285, A_distance=1.080239; B_neighbor= 285, B_distance=1.080186; neighbor_gate_cos=+0.577817, neighbor_up_cos=+0.062385; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L04 unit=1804: A_neighbor=1106, A_distance=1.064846; B_neighbor=1106, B_distance=1.064860; neighbor_gate_cos=+0.433604, neighbor_up_cos=-0.298068; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L04 unit=1044: A_neighbor= 316, A_distance=1.135266; B_neighbor= 316, B_distance=1.135236; neighbor_gate_cos=+0.506256, neighbor_up_cos=+0.036097; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L05 unit= 584: A_neighbor= 200, A_distance=1.039044; B_neighbor= 200, B_distance=1.039761; neighbor_gate_cos=+0.460536, neighbor_up_cos=+0.134748; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L06 unit=1824: A_neighbor= 560, A_distance=1.301930; B_neighbor= 560, B_distance=1.302027; neighbor_gate_cos=+0.164336, neighbor_up_cos=-0.033414; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L10 unit=1363: A_neighbor=1106, A_distance=1.132240; B_neighbor=1106, B_distance=1.131844; neighbor_gate_cos=+0.552383, neighbor_up_cos=-0.663011; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L10 unit=1564: A_neighbor=1259, A_distance=1.128584; B_neighbor=1259, B_distance=1.128573; neighbor_gate_cos=+0.400091, neighbor_up_cos=+0.277541; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L10 unit=1529: A_neighbor=1012, A_distance=1.205758; B_neighbor=1012, B_distance=1.205810; neighbor_gate_cos=+0.390633, neighbor_up_cos=+0.166234; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L13 unit=  16: A_neighbor= 829, A_distance=1.006730; B_neighbor= 829, B_distance=1.006427; neighbor_gate_cos=+0.614767, neighbor_up_cos=-0.176506; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L14 unit= 641: A_neighbor=1798, A_distance=1.054291; B_neighbor=1798, B_distance=1.054393; neighbor_gate_cos=+0.568080, neighbor_up_cos=-0.124723; exact_group_sizes_A/B=1/1; inactive_A/B=False/False
L16 unit= 893: A_neighbor= 324, A_distance=1.049521; B_neighbor= 324, B_distance=1.049546; neighbor_gate_cos=+0.541711, neighbor_up_cos=-0.557632; exact_group_sizes_A/B=1/1; inactive_A/B=False/False

=== RAW VERSUS DIAGONALLY SCALED GATE SEPARATION ===
 L nearest_other_distance_median_raw_A median_scaled_A median_scaled_B
 0                            1.267896        1.228242        1.228075
 1                            1.287700        1.294016        1.294061
 2                            1.294072        1.285317        1.285181
 3                            1.231337        1.245649        1.245770
 4                            1.241456        1.189822        1.189892
 5                            1.213284        1.184118        1.184129
 6                            1.241415        1.203177        1.203003
 7                            1.249083        1.201557        1.201721
 8                            1.228391        1.173372        1.173335
 9                            1.241038        1.148259        1.148500
10                            1.234329        1.087252        1.086539
11                            1.225529        1.067024        1.066967
12                            1.249578        1.060294        1.060619
13                            1.286203        1.073557        1.073081
14                            1.288849        1.061193        1.061347
15                            1.289908        1.063589        1.063370
16                            1.290656        1.054416        1.054660
17                            1.268821        1.048296        1.048482

=== EXACT REDUNDANCY AND CORRESPONDENCE SUMMARY ===
Total same-index A-to-B nearest gates: 36864 of 36864
Total same-index B-to-A nearest gates: 36864 of 36864

Elapsed seconds: 7.57

Interpretation limits:
Unique gate correspondence rules out a simple unit-permutation explanation.
Absence of exact duplicate gates rules out that particular mixing symmetry,
not all possible functional redundancies or compensations.
Nearness of weight rows does not establish coactivation on real prompts.
No gate, activation, hidden-state normalization, or model input was evaluated.

```