OUTPUT of run 21 of at most 60:
```
EMBEDDING-SELECTED NEIGHBOR PAIRS
Pool size: 4096 pair count: 512
These are geometric neighbors, NOT factual question/answer pairs.
Neighbor cosine quantiles: [0.1412 0.2287 0.2823 0.3537 0.7151]
' therapists' <-> ' consultants' cos=0.368
' Results' <-> ' findings' cos=0.404
' ANSI' <-> ' HTTP' cos=0.234
' particular' <-> ' speci' cos=0.415
' compartir' <-> ' partage' cos=0.519
' sheriff' <-> ' detective' cos=0.303
' vielleicht' <-> ' pourrait' cos=0.306
' operasi' <-> ' funcionamiento' cos=0.191
' Eagles' <-> ' Steelers' cos=0.325
' backup' <-> ' storage' cos=0.291
' Brasil' <-> ' Thailand' cos=0.459
' Kung' <-> ' Cheng' cos=0.182
' terrifying' <-> ' scary' cos=0.688
' render' <-> ' make' cos=0.294
' harmony' <-> ' balance' cos=0.283
' Geometric' <-> ' Mathematical' cos=0.373
' SCRE' <-> ' sque' cos=0.310
' Desde' <-> ' sejak' cos=0.351
' lung' <-> ' Parkinson' cos=0.226
' maggiore' <-> ' higher' cos=0.296

TEMPLATE COSINES, FIRST THREE
[[1.     0.6402 0.7079]
 [0.6402 1.     0.8273]
 [0.7079 0.8273 1.    ]]

ATTENTION IDENTITY-DISTINCTION DIFFERENCES, B-A *10^4
Columns: ['fine_neighbor_identity', 'coarse_pair_features', 'random_pair_identity', 'fine_half_1', 'fine_half_2']
L00H2   :   +0.1016   +0.1023   +0.0743   +0.1213   +0.0819
L10H0   :   +0.0748   +0.0742   +0.0819   +0.0658   +0.0838
L10H2   :   +0.0704   -0.0019   +0.0107   +0.0614   +0.0795
L11H1   :   +0.0620   -0.0037   +0.0157   +0.1069   +0.0171
L07H0   :   +0.0561   +0.0379   +0.0642   +0.1068   +0.0053
L04H1   :   +0.0518   +0.0612   +0.0642   +0.0504   +0.0531
L00H1   :   -0.1598   -0.1505   -0.1574   -0.2102   -0.1094
L17H1   :   -0.1480   -0.1740   -0.2017   -0.1665   -0.1296
L02H0   :   -0.1069   -0.0915   -0.0831   -0.1136   -0.1003
L03H3   :   -0.0999   -0.0106   -0.0687   -0.1048   -0.0950
L16H3   :   -0.0719   -0.0853   -0.1210   -0.0601   -0.0838
L02H3   :   -0.0682   -0.0626   -0.1062   -0.0820   -0.0543
L00H3   :   +0.0073   +0.0406   -0.0063   +0.0259   -0.0113
L05H0   :   -0.0125   -0.0022   +0.0011   -0.0050   -0.0200
L17H2   :   -0.0245   -0.0317   -0.0469   -0.0022   -0.0469

ALL MLP UNGATED IDENTITY-DISTINCTION DIFFERENCES, B-A *10^4
These omit MLP gate coefficients.
L00MLP_ungated :   +0.2651   +0.1698   +0.1833   +0.3166   +0.2136
L01MLP_ungated :   +0.0574   +0.0121   -0.0717   +0.1183   -0.0035
L02MLP_ungated :   -0.0824   -0.0506   -0.0323   -0.0830   -0.0818
L03MLP_ungated :   +0.0213   -0.0065   -0.0240   +0.0558   -0.0132
L04MLP_ungated :   -0.0101   +0.0218   +0.0119   -0.0052   -0.0149
L05MLP_ungated :   +0.0238   +0.0428   +0.0451   -0.0057   +0.0534
L06MLP_ungated :   -0.0181   +0.0145   +0.0062   -0.0255   -0.0107
L07MLP_ungated :   +0.0346   +0.0404   +0.0543   +0.0370   +0.0322
L08MLP_ungated :   -0.0253   -0.0234   -0.0040   -0.0343   -0.0164
L09MLP_ungated :   -0.0004   +0.0069   +0.0152   +0.0197   -0.0204
L10MLP_ungated :   +0.0304   -0.0033   -0.0209   +0.0132   +0.0476
L11MLP_ungated :   -0.0314   -0.0288   -0.0327   -0.0309   -0.0318
L12MLP_ungated :   -0.0072   +0.0167   +0.0308   +0.0152   -0.0295
L13MLP_ungated :   -0.0364   -0.0315   -0.0132   -0.0248   -0.0479
L14MLP_ungated :   -0.1094   -0.0517   -0.0789   -0.1064   -0.1125
L15MLP_ungated :   +0.1184   +0.1291   +0.1253   +0.1336   +0.1033
L16MLP_ungated :   -0.0122   +0.0338   +0.0304   +0.0112   -0.0356
L17MLP_ungated :   +0.2778   +0.1227   +0.1246   +0.2339   +0.3216

DESCRIPTIVE CONSISTENCY COUNTS
attention fine-identity B-positive: 39 / 72 positive in both halves: 25 negative in both halves: 23
mlp_ungated fine-identity B-positive: 8 / 18 positive in both halves: 5 negative in both halves: 7
Neither components nor vocabulary pairs are statistically independent.

WHICH NEIGHBOR DISTINCTIONS DRIVE SELECTED CHANGES?

 L00H1 baseline A/B fine coupling *1e4: -119.59844545219433 -119.75810492370414
largest B-favoring [("' Alumni'", "' graduates'", 4.0114), ("' make'", "' Making'", 3.3781), ("' EURO'", "' Ireland'", 3.0533), ("' wrench'", "' Crack'", 2.9223)]
largest A-favoring [("' segundos'", "' Second'", -3.5312), ("' ensino'", "' Teacher'", -3.3634), ("' haired'", "' blond'", -3.1716), ("' discourage'", "' prevent'", -3.0412)]

 L17H1 baseline A/B fine coupling *1e4: 273.5303609655588 273.38188395493614
largest B-favoring [("' EURO'", "' Ireland'", 1.8832), ("' median'", "' marginal'", 1.7951), ("' Content'", "' Music'", 1.7687), ("' nights'", "' Days'", 1.4814)]
largest A-favoring [("' Brasil'", "' Thailand'", -3.2678), ("' appara'", "' machine'", -3.1079), ("' reach'", "' alcance'", -2.2339), ("' ensino'", "' Teacher'", -2.1658)]

 L17H2 baseline A/B fine coupling *1e4: -50.36032913209436 -50.384701744650215
largest B-favoring [("' jaune'", "' turquoise'", 1.0847), ("' nights'", "' Days'", 1.0285), ("' CONSTRAINT'", "' LinearLayout'", 0.9813), ("' Kreis'", "' Riemann'", 0.8658)]
largest A-favoring [("' process'", "' Workflow'", -1.2109), ("' Brasil'", "' Thailand'", -1.0739), ("' advert'", "' publicidad'", -1.0034), ("' Lennon'", "' Dylan'", -0.9881)]

Only weight matrices, embedding-row geometry, and their linear products were used.
No text sequence, gate, activation function, state normalization, or model output was evaluated.

```