OUTPUT of run 10 of at most 60:
```
=== INPUT-PROJECTION DIFFERENCE SUBSPACES ===
Vocabulary rows examined: 150357
Scores are squared cosine projections of embedding rows onto weight
difference subspaces; they are not token probabilities or activations.
undo_input_diagonal divides matrix columns by learned input diagonals.
It does not normalise an embedding or hidden state.
For an ordinary gradient this can undo a known linear input scaling;
the unknown optimiser may invalidate that gradient-footprint interpretation.
Later-layer rows are included only as comparisons, not direct token recovery.

L0 qkv raw captured difference energy: top8=45.63%, top16=57.76%
 span8:  '\n':0.777, 'What':0.703, ' Argentina':0.682, '1':0.676, ' Brazil':0.671, ' Brasil':0.654, ' ':0.618, ' Italy':0.593, ' What':0.592, ' Colombia':0.589, ' Chile':0.588, ' Poland':0.572
 span16: '\n':0.800, ' minus':0.742, 'What':0.720, ' Brasil':0.719, ' is':0.711, ' Argentina':0.710, ' Brazil':0.700, '1':0.684, ' ':0.684, ' BLUENRG':0.633, ' MDLVertex':0.629, ' gstlal':0.628
 mode0 +[' Giveen':+0.426, ' Tjiwarl':+0.413, '<unused185>':+0.406] -[' ':-0.573, ' is':-0.569, ' super':-0.551]
 mode1 +[' Argentina':+0.734, ' Brasil':+0.691, 'Argentina':+0.690] -['\n':-0.235, ' ':-0.192, ' in':-0.158]

L0 qkv undo_input_diagonal captured difference energy: top8=45.67%, top16=57.91%
 span8:  '\n':0.750, 'What':0.682, '1':0.622, ' Argentina':0.615, ' Brazil':0.600, ' Brasil':0.583, ' What':0.569, ' ':0.553, ' Italy':0.541, ' Poland':0.535, ' Chile':0.529, ' Japan':0.529
 span16: '\n':0.777, 'What':0.707, ' minus':0.683, ' is':0.665, ' Brasil':0.645, ' Brazil':0.633, '1':0.632, ' Argentina':0.631, ' ':0.619, ' BLUENRG':0.619, ' MDLVertex':0.610, ' gstlal':0.609
 mode0 +[' Giveen':+0.494, ' BLUENRG':+0.481, ' gstlal':+0.477] -[' ':-0.536, ' is':-0.535, '\n':-0.501]
 mode1 +[' Argentina':+0.699, ' Brazil':+0.680, ' Brasil':+0.663] -['\n':-0.075, 'hatian':-0.073, ' tarik':-0.071]

L0 q undo_input_diagonal captured difference energy: top8=44.53%, top16=59.12%
 span8:  ' Argentina':0.634, ' minus':0.627, ' Brazil':0.594, '1':0.588, ' Italy':0.581, ' Poland':0.574, ' Germany':0.573, ' BLUENRG':0.569, ' France':0.566, ' India':0.559, ' Spain':0.558, ' CCOc':0.557
 span16: 'ia':0.786, 'model':0.736, ' minus':0.735, ' pumpkin':0.722, ' eggplant':0.672, ' Argentina':0.669, 'user':0.655, ' plus':0.636, ' Brazil':0.618, '7':0.610, ' BLUENRG':0.609, '<start_of_turn>':0.603
 mode0 +['1':+0.099, 'elij':+0.065, 'intas':+0.064] -[' Argentina':-0.731, ' Colombia':-0.699, ' Brazil':-0.695]
 mode1 +['\n':+0.582, ' ':+0.485, ' and':+0.482] -[' MDLVertex':-0.619, ' BLUENRG':-0.615, ' gstlal':-0.612]

L0 k undo_input_diagonal captured difference energy: top8=55.28%, top16=67.52%
 span8:  '\n':0.623, 'What':0.613, ' Brasil':0.577, ' is':0.565, ' Argentina':0.535, ' Brazil':0.524, ' What':0.502, '1':0.489, ' Japan':0.468, ' Indonesia':0.467, ' Italy':0.465, ' Germany':0.455
 span16: '\n':0.699, ' Brasil':0.661, 'What':0.644, ' is':0.644, 'user':0.611, ' color':0.605, ' Argentina':0.596, ' Brazil':0.589, ' I':0.583, '1':0.581, 'model':0.576, ' continent':0.543
 mode0 +[' Giveen':+0.402, ' Vestone':+0.391, ' Tjiwarl':+0.385] -[' is':-0.503, ' has':-0.411, ' in':-0.394]
 mode1 +['<bos>':+0.302, ' sacrificio':+0.200, ' visitante':+0.199] -[' India':-0.546, ' Japan':-0.544, ' Canada':-0.543]

L0 v undo_input_diagonal captured difference energy: top8=61.41%, top16=71.29%
 span8:  '\n':0.732, 'What':0.670, ' ':0.602, '1':0.569, ' What':0.555, ' Brasil':0.521, 'what':0.506, ' what':0.502, 'The':0.501, 'WHAT':0.489, ' Brazil':0.480, ' The':0.473
 span16: '\n':0.744, 'What':0.680, " '":0.676, ' United':0.648, ' Brasil':0.616, 'Which':0.615, 'How':0.606, ' ':0.606, ' is':0.604, '1':0.593, ' What':0.579, 'The':0.566
 mode0 +['\n':+0.661, '\n\n':+0.500, '\n\n\n':+0.483] -['<unused1832>':-0.249, '<unused237>':-0.242, '<unused1118>':-0.233]
 mode1 +[' ':+0.569, ' in':+0.388, ' a':+0.375] -[' gstlal':-0.488, ' BSData':-0.483, ' RUTARE':-0.483]

L1 qkv undo_input_diagonal captured difference energy: top8=69.55%, top16=79.68%
 span8:  ' Brasil':0.162, 's':0.144, 'Brasil':0.138, ' Brazil':0.116, ' Canada':0.110, 'Canada':0.109, 'StdHandle':0.108, 'Indonesia':0.108, ' is':0.107, ' Mexico':0.107, ' Japan':0.107, ' Australia':0.105
 span16: 's':0.216, ' minus':0.199, 'model':0.188, ' Brasil':0.177, 'what':0.174, 'Brasil':0.171, 'What':0.163, ' model':0.156, '6':0.153, '7':0.150, 'Canada':0.148, ' what':0.146
 mode0 +['setPrototypeOf':+0.195, 'StdHandle':+0.194, '\n':+0.174] -['dropna':-0.191, 'gulation':-0.184, '="#">':-0.180]
 mode1 +['7':+0.235, '8':+0.224, '1':+0.222] -['IODevice':-0.189, 'utrients':-0.163, ' CCc':-0.158]

L2 qkv undo_input_diagonal captured difference energy: top8=69.93%, top16=83.21%
 span8:  '<tfoot>':0.066, 'oresizing':0.065, 'maxcdn':0.064, 'nlm':0.064, 'utrients':0.062, 'znych':0.061, 'tio':0.061, 'labelledby':0.060, 'stackpath':0.059, 'ibacter':0.058, 'gulation':0.057, 'TestSuite':0.057
 span16: 'LEncoder':0.102, 'CompilerServices':0.097, 's':0.095, 'tableFuture':0.092, 'e':0.090, 'mentare':0.090, 'tir':0.088, 'ernetes':0.086, ' adipisicing':0.086, 'tio':0.085, 'IODevice':0.084, 'ojs':0.084
 mode0 +['setPrototypeOf':+0.195, 'TestSuite':+0.155, 'roweak':+0.151] -['<tfoot>':-0.227, 'dni':-0.205, ' []:':-0.202]
 mode1 +['adrat':+0.197, ' taea':+0.166, 'mentare':+0.148] -['leza':-0.172, 'omerang':-0.154, 'leys':-0.153]

L3 qkv undo_input_diagonal captured difference energy: top8=65.88%, top16=80.10%
 span8:  'CompilerServices':0.085, 'fillRect':0.076, 'stackpath':0.073, '<tfoot>':0.073, 'maxcdn':0.070, 'LEncoder':0.069, 'ilberger':0.069, 'jsce':0.066, 'seekg':0.063, 'icletas':0.063, ' []:':0.063, 'illerato':0.063
 span16: 'CompilerServices':0.149, 'e':0.119, 'LEncoder':0.115, 'fillRect':0.113, 'PushMatrix':0.106, 'sman':0.103, 'Hochspringen':0.103, '<unused1640>':0.102, '="#">':0.101, 'tableFuture':0.100, 'labelledby':0.100, 's':0.098
 mode0 +['setPrototypeOf':+0.172, 'itriangular':+0.170, 'vegarde':+0.147] -['<tfoot>':-0.225, 'dni':-0.211, 'stackpath':-0.207]
 mode1 +['uluh':+0.135, 'agascar':+0.133, 'ubert':+0.127] -['argsort':-0.156, 'hrlich':-0.151, 'codile':-0.143]

=== L0 JOINT SPAN: EXCESS OVER SIGN-SCRAMBLED WEIGHT CONTROLS ===
Top 30 tokens; four controls are a diagnostic, not a calibrated null test.
'\n':0.754, 'What':0.673, ' minus':0.653, ' is':0.649, ' Brasil':0.623, ' Brazil':0.615, ' Argentina':0.611, '1':0.611, ' ':0.600, ' BLUENRG':0.574, ' What':0.567, ' gstlal':0.565, ' MDLVertex':0.562, ' BSData':0.559, '<end_of_turn>':0.558, ' CCOc':0.557, ' Pusenkoff':0.557, ' yojet':0.554, ' saddas':0.550, ' dissanti':0.549, 'labeltree':0.548, ' kahelaro':0.548, ' BlackSholes':0.548, '<start_of_turn>':0.545, 'jgamma':0.545, ' SPARSEBUNDLE':0.545, ' vattabbam':0.544, ' vacanato':0.544, 'positroids':0.544, ' Italy':0.543

=== FIXED TOKEN FAMILIES: OBSERVED / NULL SPAN16 PROJECTION ===
digits [('0', 0.382, 0.02), ('1', 0.632, 0.022), ('2', 0.562, 0.024), ('3', 0.537, 0.023), ('4', 0.493, 0.022), ('5', 0.457, 0.021), ('6', 0.457, 0.02), ('7', 0.489, 0.019), ('8', 0.469, 0.018), ('9', 0.495, 0.021)]
belief [('think', 0.067, 0.024), (' think', 0.154, 0.029), ('believe', 0.079, 0.021), (' believe', 0.125, 0.023), ('guess', 0.093, 0.031), (' guess', 0.098, 0.027), (' suppose', 0.1, 0.032), ('opinion', 0.049, 0.029), (' opinion', 0.171, 0.033), ('assume', 0.106, 0.024), (' assume', 0.087, 0.023)]
arithmetic [('plus', 0.296, 0.032), (' plus', 0.492, 0.043), ('minus', 0.349, 0.032), (' minus', 0.683, 0.03), ('times', 0.178, 0.03), (' times', 0.376, 0.034), ('multiply', 0.133, 0.026), (' multiply', 0.164, 0.03), ('divide', 0.116, 0.026), (' divide', 0.177, 0.029), ('equals', 0.094, 0.034), (' equals', 0.203, 0.02)]
question_answer [('what', 0.539, 0.03), (' what', 0.548, 0.03), ('question', 0.145, 0.027), (' question', 0.201, 0.036), ('answer', 0.109, 0.02), (' answer', 0.188, 0.028), ('result', 0.078, 0.027), (' result', 0.141, 0.03), ('correct', 0.072, 0.031), (' correct', 0.191, 0.035)]
chat [('user', 0.552, 0.029), (' user', 0.373, 0.034), ('model', 0.375, 0.017), (' model', 0.365, 0.015), ('assistant', 0.078, 0.025), (' assistant', 0.17, 0.027), ('<start_of_turn>', 0.599, 0.054), ('<end_of_turn>', 0.596, 0.037)]

```