OUTPUT of run 8 of at most 20:
```
Descriptive dictionary: 53999 space-prefixed lowercase tokens
belief single-token anchors: [' think', ' guess', ' believe', ' opinion', ' suppose']
evidence single-token anchors: [' fact', ' evidence', ' accurate', ' correct', ' actual']
question single-token anchors: [' what', ' which', ' question', ' why', ' how']
answer single-token anchors: [' answer', ' response', ' reply', ' solution', ' result']

=== JOINT MLP GATE/UP INPUT-CHANGE SUBSPACES ===
Stacked delta gate and delta up matrices, including only diagonal pre-MLP scales.
Vocabulary poles describe changed read directions, NOT which units activate.
Capture = fraction of a lexical contrast lying in the top-six changed-read subspace.
L00 energy1=0.114 energy6=0.346 units=[329, 16, 1536, 281, 1305] capture={'belief-evidence': 0.0179, 'answer-question': 0.0179}
 + [' and', ' encima', ' uprav', ' tratto', ' inmedi']  - [' russia', ' japan', ' korea', ' germany', ' africa']
L01 energy1=0.064 energy6=0.262 units=[552, 273, 1466, 1269, 1371] capture={'belief-evidence': 0.0117, 'answer-question': 0.0218}
 + [' pih', ' vab', ' encima', ' incum', ' ihtiy']  - [' korea', ' japan', ' russia', ' korean', ' germany']
L02 energy1=0.074 energy6=0.260 units=[1737, 1335, 594, 671, 1683] capture={'belief-evidence': 0.0058, 'answer-question': 0.0026}
 + [' scop', ' sebagaimana', ' spirito', ' averiguar', ' conosci']  - [' coaching', ' minus', ' paris', ' darken', ' drilling']
L03 energy1=0.048 energy6=0.237 units=[700, 338, 1140, 254, 1890] capture={'belief-evidence': 0.0032, 'answer-question': 0.0068}
 + [' vasos', ' camiseta', ' relevancia', ' assistir', ' invitado']  - [' discrete', ' sleeping', ' compact', ' four', ' drilling']
L04 energy1=0.095 energy6=0.307 units=[1193, 1885, 1787, 282, 776] capture={'belief-evidence': 0.009, 'answer-question': 0.0057}
 + [' banks', ' vertically', ' simultaneous', ' vertical', ' scientist']  - [' proporcionar', ' incorporar', ' oferecer', ' localizar', ' trover']
L05 energy1=0.069 energy6=0.247 units=[897, 1568, 1392, 447, 1035] capture={'belief-evidence': 0.004, 'answer-question': 0.0109}
 + [' denominador', ' incorporar', ' queridos', ' menikmati', ' cuyos']  - [' woman', ' equality', ' hyp', ' orthodox', ' textbooks']
L06 energy1=0.085 energy6=0.275 units=[1703, 117, 501, 1522, 187] capture={'belief-evidence': 0.0096, 'answer-question': 0.0094}
 + [' locales', ' poblaciones', ' propia', ' vostri', ' popula']  - [' sata', ' banana', ' supermarket', ' python', ' stereo']
L07 energy1=0.108 energy6=0.327 units=[186, 256, 984, 1022, 775] capture={'belief-evidence': 0.0037, 'answer-question': 0.0062}
 + [' pizzas', ' sata', ' electors', ' adenine', ' plaques']  - [' ausgel', ' pengendalian', ' hass', ' moviment', ' lingk']
L08 energy1=0.123 energy6=0.334 units=[898, 454, 1313, 7, 659] capture={'belief-evidence': 0.0058, 'answer-question': 0.0041}
 + [' dva', ' meters', ' deductions', ' dicas', ' trzy']  - [' conjunt', ' tailoring', ' lingk', ' snuff', ' exertions']
L09 energy1=0.123 energy6=0.359 units=[1950, 1213, 774, 647, 1635] capture={'belief-evidence': 0.0037, 'answer-question': 0.0027}
 + [' irradi', ' sakura', ' tailoring', ' tho', ' quartier']  - [' incorrect', ' excitatory', ' three', ' dva', ' erroneous']
L10 energy1=0.155 energy6=0.427 units=[196, 91, 816, 362, 1802] capture={'belief-evidence': 0.0023, 'answer-question': 0.0056}
 + [' plantations', ' irradi', ' madame', ' preferentially', ' multin']  - [' trei', ' nueve', ' veces', ' patru', ' sevent']
L11 energy1=0.146 energy6=0.401 units=[954, 1631, 691, 406, 1266] capture={'belief-evidence': 0.0052, 'answer-question': 0.0062}
 + [' eight', ' vijf', ' vier', ' hundred', ' nueve']  - [' istanbul', ' bilayers', ' zindagi', ' plantations', ' customise']
L12 energy1=0.143 energy6=0.413 units=[146, 1870, 888, 385, 280] capture={'belief-evidence': 0.0028, 'answer-question': 0.0045}
 + [' zehn', ' sechs', ' sieben', ' eight', ' neun']  - [' zindagi', ' toot', ' acclimat', ' turbulence', ' mohabbat']
L13 energy1=0.172 energy6=0.449 units=[354, 1885, 1679, 1128, 1199] capture={'belief-evidence': 0.0089, 'answer-question': 0.0047}
 + [' exfol', ' convers', ' railing', ' gratitude', ' marm']  - [' sieben', ' zehn', ' eight', ' sechs', ' oito']
L14 energy1=0.170 energy6=0.430 units=[1032, 393, 219, 881, 1779] capture={'belief-evidence': 0.0016, 'answer-question': 0.0035}
 + [' retraining', ' plages', ' spieg', ' pptn', ' mts']  - [' sieben', ' eight', ' sechs', ' nine', ' playable']
L15 energy1=0.140 energy6=0.389 units=[422, 51, 1399, 2036, 1183] capture={'belief-evidence': 0.0033, 'answer-question': 0.0054}
 + [' sip', ' improv', ' bleed', ' bourbon', ' ska']  - [' sieben', ' sechs', ' vogliamo', ' oito', ' patru']
L16 energy1=0.136 energy6=0.387 units=[1633, 206, 714, 829, 503] capture={'belief-evidence': 0.0019, 'answer-question': 0.0054}
 + [' sieben', ' sechs', ' zehn', ' patru', ' eight']  - [' fasse', ' tume', ' avail', ' kabhi', ' stables']
L17 energy1=0.130 energy6=0.370 units=[585, 108, 301, 1172, 1942] capture={'belief-evidence': 0.0036, 'answer-question': 0.0076}
 + [' wcex', ' viewport', ' dissati', ' datasource', ' whisky']  - [' hundred', ' sieben', ' nine', ' patru', ' oito']

=== EARLIER MLP OUTPUT-CHANGE CONNECTIVITY TO CANDIDATE Q/K MAPS ===
gain = ||R deltaD||F / ||deltaD||F, with ||R||F=1.
isotropic reference gain = 1/sqrt(640) = 0.03953
Enrichment is structural alignment, NOT causal importance.
L00 mean_gain=0.00805 enrichment=0.204 targets=[(9, 2, np.float64(0.00549)), (10, 2, np.float64(0.00567)), (11, 2, np.float64(0.00782)), (12, 1, np.float64(0.01059)), (12, 2, np.float64(0.01067))]
 top connected changed-output units: [(669, 0.00205), (1618, 0.00205), (220, 0.00202), (273, 0.00185), (540, 0.00183), (840, 0.00178)]
L01 mean_gain=0.00663 enrichment=0.168 targets=[(9, 2, np.float64(0.00481)), (10, 2, np.float64(0.005)), (11, 2, np.float64(0.00666)), (12, 1, np.float64(0.00833)), (12, 2, np.float64(0.00835))]
 top connected changed-output units: [(306, 0.00182), (846, 0.00149), (827, 0.00147), (463, 0.00145), (375, 0.00137), (1487, 0.00131)]
L02 mean_gain=0.00780 enrichment=0.197 targets=[(9, 2, np.float64(0.00593)), (10, 2, np.float64(0.00617)), (11, 2, np.float64(0.00802)), (12, 1, np.float64(0.00943)), (12, 2, np.float64(0.00947))]
 top connected changed-output units: [(1167, 0.00141), (2046, 0.00138), (1019, 0.00132), (105, 0.00131), (1470, 0.00125), (1308, 0.00125)]
L03 mean_gain=0.00697 enrichment=0.176 targets=[(9, 2, np.float64(0.00512)), (10, 2, np.float64(0.00536)), (11, 2, np.float64(0.008)), (12, 1, np.float64(0.00821)), (12, 2, np.float64(0.00816))]
 top connected changed-output units: [(958, 0.00146), (1394, 0.0014), (1611, 0.00134), (1088, 0.00127), (385, 0.00126), (263, 0.00123)]
L04 mean_gain=0.00503 enrichment=0.127 targets=[(9, 2, np.float64(0.005)), (10, 2, np.float64(0.00448)), (11, 2, np.float64(0.00512)), (12, 1, np.float64(0.00527)), (12, 2, np.float64(0.00528))]
 top connected changed-output units: [(1387, 0.00134), (1718, 0.00134), (2004, 0.0013), (119, 0.00121), (209, 0.00121), (445, 0.00116)]
L05 mean_gain=0.00676 enrichment=0.171 targets=[(9, 2, np.float64(0.00726)), (10, 2, np.float64(0.00652)), (11, 2, np.float64(0.00679)), (12, 1, np.float64(0.00661)), (12, 2, np.float64(0.00663))]
 top connected changed-output units: [(389, 0.00175), (1714, 0.00158), (1304, 0.00147), (1225, 0.00145), (947, 0.00145), (1877, 0.00144)]
L06 mean_gain=0.00895 enrichment=0.227 targets=[(9, 2, np.float64(0.01073)), (10, 2, np.float64(0.0092)), (11, 2, np.float64(0.0087)), (12, 1, np.float64(0.00807)), (12, 2, np.float64(0.00806))]
 top connected changed-output units: [(887, 0.00118), (1898, 0.00107), (1449, 0.00105), (546, 0.00104), (1184, 0.00103), (1925, 0.00098)]
L07 mean_gain=0.01099 enrichment=0.278 targets=[(9, 2, np.float64(0.0146)), (10, 2, np.float64(0.01161)), (11, 2, np.float64(0.01023)), (12, 1, np.float64(0.00925)), (12, 2, np.float64(0.00925))]
 top connected changed-output units: [(391, 0.00165), (1140, 0.00106), (1607, 0.00099), (691, 0.00097), (947, 0.00097), (1312, 0.00096)]
L08 mean_gain=0.01259 enrichment=0.319 targets=[(9, 2, np.float64(0.01772)), (10, 2, np.float64(0.01394)), (11, 2, np.float64(0.01141)), (12, 1, np.float64(0.00994)), (12, 2, np.float64(0.00994))]
 top connected changed-output units: [(500, 0.00129), (1098, 0.00122), (1326, 0.00109), (566, 0.00104), (778, 0.00104), (707, 0.001)]
L09 mean_gain=0.01493 enrichment=0.378 targets=[(10, 2, np.float64(0.01961)), (11, 2, np.float64(0.01516)), (12, 1, np.float64(0.01243)), (12, 2, np.float64(0.01252))]
 top connected changed-output units: [(1738, 0.00108), (1135, 0.00105), (320, 0.00097), (2030, 0.00096), (1318, 0.00094), (319, 0.00094)]
L10 mean_gain=0.01468 enrichment=0.371 targets=[(11, 2, np.float64(0.01734)), (12, 1, np.float64(0.01335)), (12, 2, np.float64(0.01335))]
 top connected changed-output units: [(1103, 0.00187), (1593, 0.00174), (774, 0.00139), (1901, 0.00136), (1896, 0.00134), (1286, 0.00131)]
L11 mean_gain=0.01315 enrichment=0.333 targets=[(12, 1, np.float64(0.01323)), (12, 2, np.float64(0.01308))]
 top connected changed-output units: [(1510, 0.00198), (1949, 0.0018), (1456, 0.00177), (1496, 0.00162), (216, 0.0016), (94, 0.00145)]

=== HIGHEST CONNECTIVITY CANDIDATES: BASELINE READ/WRITE GEOMETRY ===
L9 unit1738
  gate-read + [' restoran', ' flaps', ' ristor', ' sekret'] - [' chase', ' skinny', ' polymeric', ' ticker']
  up-read + [' mooi', ' centr', ' embryo', ' fft'] - [' scooters', ' bikes', ' toppings', ' privileges']
  output-write + [' multi', ' itinerary', ' prioritizing', ' hosting'] - [' della', ' dans', ' dell', ' nella']
L9 unit1135
  gate-read + [' panjang', ' gian', ' kort', ' cia'] - [' conno', ' anthropologists', ' milled', ' adolescents']
  up-read + [' aprovechar', ' peraturan', ' kadang', ' entrare'] - [' agave', ' rejuven', ' sque', ' jogging']
  output-write + [' hark', ' submer', ' frontline', ' stretching'] - [' kaha', ' chiuso', ' cerrado', ' fiducia']
L10 unit1103
  gate-read + [' attaques', ' versi', ' attaque', ' edizione'] - [' staunch', ' trace', ' garments', ' scarf']
  up-read + [' ming', ' juan', ' marshmallow', ' metam'] - [' depict', ' stealth', ' interiors', ' bays']
  output-write + [' mettre', ' cuatro', ' karotiti', ' cinque'] - [' multi', ' representation', ' pre', ' especially']
L10 unit1593
  gate-read + [' czyn', ' kritis', ' histidine', ' essentielle'] - [' tunneling', ' hairstyles', ' dreams', ' www']
  up-read + [' pourcentage', ' actives', ' radiance', ' porcentaje'] - [' valmist', ' recoil', ' tweezers', ' refractory']
  output-write + [' hacer', ' mettre', ' meny', ' realizar'] - [' multi', ' representation', ' represent', ' swallow']
L11 unit1510
  gate-read + [' heridas', ' occidentale', ' incidente', ' ejecutar'] - [' whale', ' surfers', ' conformity', ' whales']
  up-read + [' folly', ' genius', ' vantagem', ' storey'] - [' molte', ' unfortunately', ' malheureusement', ' contrairement']
  output-write + [' natu', ' publish', ' restore', ' flor'] - [' wur', ' runners', ' christian', ' ederim']
L11 unit1949
  gate-read + [' gsap', ' ignite', ' gleaned', ' konkret'] - [' outer', ' feedings', ' oreilles', ' goodness']
  up-read + [' ases', ' rokov', ' jaye', ' asesor'] - [' hostname', ' multimeter', ' alternator', ' lumen']
  output-write + [' txtbtn', ' hacer', ' realizar', ' moet'] - [' cass', ' recon', ' def', ' dam']

All MLP analyses treated matrices separately or composed linear parameter maps.

```