=== OUTPUT-CHANGE ALIGNMENT WITH NINE-DIMENSIONAL TOKEN SUBSPACES ===
Each entry is the fraction of squared output-weight difference captured
by an orthogonal embedding-contrast subspace, not behavioural importance.
Controls: 32 groups of ten vocabulary rows, each yielding nine contrasts.
L geo | digits% words% labels% | control mean% sd% | digit_z word_z
 0 raw  |   1.65   1.45   1.33 |  1.35  0.07 |  +4.38  +1.51
 0 diag |   0.84   1.18   2.81 |  0.97  0.13 |  -0.97  +1.63
 1 raw  |   1.97   1.48   1.52 |  1.40  0.06 | +10.30  +1.53
 1 diag |   0.62   1.33   1.23 |  0.59  0.10 |  +0.30  +7.69
 2 raw  |   2.24   1.73   1.38 |  1.43  0.09 |  +9.04  +3.28
 2 diag |   1.35   1.33   1.66 |  1.08  0.09 |  +2.90  +2.68
 3 raw  |   1.26   1.51   1.52 |  1.41  0.09 |  -1.69  +1.01
 3 diag |   1.22   1.24   1.51 |  1.11  0.09 |  +1.26  +1.51
 4 raw  |   1.29   1.32   1.42 |  1.39  0.11 |  -0.91  -0.59
 4 diag |   1.07   1.43   1.36 |  1.26  0.11 |  -1.71  +1.46
 5 raw  |   1.40   1.38   1.39 |  1.39  0.08 |  +0.13  -0.17
 5 diag |   0.86   1.34   1.13 |  1.14  0.09 |  -3.19  +2.16
 6 raw  |   1.39   1.30   1.35 |  1.41  0.09 |  -0.22  -1.14
 6 diag |   0.94   1.40   1.02 |  1.22  0.09 |  -3.20  +2.07
 7 raw  |   1.51   1.50   1.46 |  1.41  0.09 |  +1.09  +1.06
 7 diag |   0.92   1.55   1.16 |  1.27  0.07 |  -5.14  +4.15
 8 raw  |   1.44   1.42   1.42 |  1.40  0.10 |  +0.39  +0.19
 8 diag |   0.94   1.51   1.08 |  1.26  0.09 |  -3.78  +2.81
 9 raw  |   1.52   1.50   1.39 |  1.37  0.09 |  +1.54  +1.33
 9 diag |   0.93   1.52   1.03 |  1.24  0.09 |  -3.23  +2.95
10 raw  |   1.51   1.46   1.36 |  1.38  0.09 |  +1.40  +0.86
10 diag |   0.99   1.40   0.98 |  1.22  0.10 |  -2.41  +1.93
11 raw  |   2.03   1.65   1.57 |  1.44  0.11 |  +5.19  +1.89
11 diag |   2.48   1.65   1.21 |  1.14  0.08 | +16.27  +6.26
12 raw  |   3.07   1.97   1.46 |  1.50  0.10 | +15.25  +4.59
12 diag |   3.98   1.82   1.21 |  1.17  0.09 | +32.37  +7.43
13 raw  |   4.35   2.33   1.51 |  1.53  0.10 | +29.32  +8.36
13 diag |   5.26   2.09   1.22 |  1.24  0.07 | +55.40 +11.61
14 raw  |   8.64   3.03   1.55 |  1.56  0.10 | +73.72 +15.35
14 diag |   8.19   2.72   1.32 |  1.34  0.08 | +88.49 +17.86
15 raw  |  15.76   3.68   1.51 |  1.61  0.13 | +105.23 +15.41
15 diag |  14.21   3.33   1.28 |  1.41  0.11 | +113.21 +16.92
16 raw  |  27.47   5.57   1.57 |  1.71  0.18 | +140.68 +21.11
16 diag |  24.15   5.22   1.34 |  1.56  0.18 | +127.40 +20.66
17 raw  |  40.00   8.63   2.60 |  2.16  0.29 | +132.65 +22.66
17 diag |  30.58   9.09   2.48 |  2.35  0.46 | +60.84 +14.52

Gate/up cue contrast: belief minus calculation.
Belief: ['think', 'believe', 'guess', 'assume', 'opinion', 'suppose'] Calculation: ['calculate', 'plus', 'minus', 'multiply', 'divide', 'equals']
A positive delta cosine means B's weight change points more toward
this contrast; it does NOT establish increased unit activation.

=== SELECTED UNIT WEIGHT-CHANGE SUMMARY ===
L unit | delta_gate cue cosine | delta_up cue cosine | relative output change
13  163 | -0.0339 | -0.0176 | 0.0182
13 1839 | +0.0430 | +0.0253 | 0.0195
16   36 | -0.0634 | +0.0084 | 0.0186
16 1263 | +0.0579 | +0.0204 | 0.0162
16 1125 | -0.0192 | -0.0006 | 0.0248
17 1940 | -0.0172 | +0.0166 | 0.0196
17 1516 | -0.0439 | +0.0743 | 0.0108
17 1027 | +0.0517 | +0.0356 | 0.0166
17  956 | +0.0496 | -0.0535 | 0.0277
17 1149 | -0.0065 | -0.0360 | 0.0195

=== INDIVIDUAL FACTOR VOCABULARY ASSOCIATIONS ===
Gate and up factors are inspected independently; their responses are
never multiplied. Output signs assume a positive scalar coefficient,
whose actual value or sign has NOT been computed.
L13U163 mean_gate_read: +[' UK':+0.155, ' GBP':+0.151, ' tradi':+0.150] -['izations':-0.190, ' plu':-0.181, 'appel':-0.163]
L13U163 mean_up_read: +['Ys':+0.182, ' taxonom':+0.170, 'Specification':+0.166] -[' imageData':-0.170, 'pz':-0.150, ' visible':-0.149]
L13U163 mean_output_write: +['houettes':+0.145, 'hse':+0.143, ' Profesional':+0.135] -[' Navi':-0.153, 'oot':-0.148, ' delta':-0.148]
L13U163 B-A_output_write: +[' file':+0.142, ' deep':+0.140, ' club':+0.138] -[' Seaton':-0.176, 'Rover':-0.168, 'Univ':-0.157]
L13U1839 mean_gate_read: +['otope':+0.161, ' germany':+0.154, ' Deere':+0.154] -['OUD':-0.174, ' fortuit':-0.173, 'icin':-0.164]
L13U1839 mean_up_read: +['anjutnya':+0.150, 'yank':+0.149, ' mici':+0.147] -['Diffuse':-0.189, ' irradi':-0.173, ' oxy':-0.166]
L13U1839 mean_output_write: +['umm':+0.155, 'Bump':+0.140, ' konum':+0.137] -[' FEL':-0.154, ' Redox':-0.149, ' Saxony':-0.146]
L13U1839 B-A_output_write: +[' REL':+0.157, ' redist':+0.154, ' cont':+0.153] -[' atacar':-0.148, ' accedere':-0.145, ' madeira':-0.139]
L16U36 mean_gate_read: +[' encounter':+0.177, 'margins':+0.165, ' bend':+0.151] -['idded':-0.175, 'rored':-0.172, 'atico':-0.171]
L16U36 mean_up_read: +['stock':+0.175, 'enche':+0.174, 'inplace':+0.174] -[' tujuh':-0.185, ' sizzling':-0.166, ' huit':-0.163]
L16U36 mean_output_write: +[' pengalaman':+0.197, ' ervaring':+0.195, ' EXPERIENCE':+0.192] -['acylglycer':-0.149, 'idget':-0.147, 'sType':-0.144]
L16U36 B-A_output_write: +['isiin':+0.179, ' both':+0.176, 'eduanya':+0.173] -[' tujuh':-0.246, 'Nine':-0.229, ' NORTH':-0.224]
L16U1263 mean_gate_read: +[' wastewater':+0.170, ' seafood':+0.166, ' softball':+0.165] -[' incremento':-0.158, ' pleno':-0.156, 'zai':-0.151]
L16U1263 mean_up_read: +[' fortuit':+0.170, ' depress':+0.169, ' horribly':+0.155] -['ships':-0.171, 'Cargo':-0.164, ' camadas':-0.163]
L16U1263 mean_output_write: +[' Bent':+0.150, ' ETL':+0.147, ' cilantro':+0.141] -[' buk':-0.153, ' Merr':-0.133, ' kame':-0.130]
L16U1263 B-A_output_write: +['ChooserView':+0.199, 'PICTOGRAM':+0.197, 'WholeProgram':+0.196] -['One':-0.244, 'one':-0.232, ' one':-0.230]
L16U1125 mean_gate_read: +[' kahit':+0.166, ' conformity':+0.155, ' exclusively':+0.151] -['XT':-0.164, 'ven':-0.157, 'pg':-0.155]
L16U1125 mean_up_read: +[' Up':+0.276, ' up':+0.264, 'Up':+0.243] -['queta':-0.173, 'EMPL':-0.173, 'sufficient':-0.172]
L16U1125 mean_output_write: +[' Yog':+0.134, ' jok':+0.124, 'mold':+0.124] -[' up':-0.498, 'Up':-0.450, ' Up':-0.450]
L16U1125 B-A_output_write: +[' two':+0.269, ' non':+0.258, ' self':+0.253] -['ufieurs':-0.207, 'LookAndFeels':-0.202, ' calomel':-0.192]
L17U1940 mean_gate_read: +[' aplikacji':+0.189, ' APPLICATIONS':+0.188, ' requisitos':+0.188] -[' leaving':-0.175, ' accu':-0.174, ' freed':-0.174]
L17U1940 mean_up_read: +['corners':+0.168, ' quatre':+0.164, ' strolling':+0.163] -[' MUCH':-0.178, 'Attenuation':-0.169, 'otar':-0.165]
L17U1940 mean_output_write: +[' ole':+0.138, 'otically':+0.136, 'aluronic':+0.136] -['avg':-0.133, ' decays':-0.124, ' DataLoader':-0.124]
L17U1940 B-A_output_write: +[' huit':+0.232, ' ZEN':+0.226, 'actionBarTab':+0.226] -[' in':-0.267, ' and':-0.259, ' for':-0.247]
L17U1516 mean_gate_read: +['localVarPath':+0.260, ' MILK':+0.250, 'ufieurs':+0.236] -[' to':-0.329, ' and':-0.326, ' the':-0.323]
L17U1516 mean_up_read: +['DOMstrings':+0.217, ' Lehmann':+0.206, 'ScriptAssemblies':+0.206] -[' It':-0.179, ' The':-0.176, ' with':-0.176]
L17U1516 mean_output_write: +[' PRIMARY':+0.154, ' REAL':+0.151, ' GREEN':+0.147] -[' aquifers':-0.085, ' forestry':-0.082, 'setName':-0.081]
L17U1516 B-A_output_write: +[' the':+0.330, ' The':+0.305, ' to':+0.304] -[' Atomnoj':-0.266, ' Ehnerg':-0.243, 'wallepics':-0.239]
L17U1027 mean_gate_read: +[' thirteen':+0.255, 'eight':+0.249, 'thirty':+0.236] -['otypic':-0.199, 'rinho':-0.182, 'ogenous':-0.177]
L17U1027 mean_up_read: +[' fourteen':+0.228, ' thirteen':+0.211, ' seventy':+0.208] -['justify':-0.161, 'Writable':-0.152, ' ena':-0.149]
L17U1027 mean_output_write: +['RENCY':+0.131, 'ataire':+0.131, ' prairies':+0.126] -[' FOUR':-0.293, ' FIVE':-0.282, ' cinq':-0.276]
L17U1027 B-A_output_write: +[' FOUR':+0.293, ' quatre':+0.274, ' quatro':+0.236] -[' and':-0.224, ' In':-0.223, ' in':-0.221]
L17U956 mean_gate_read: +[' fighters':+0.175, ' StreamWriter':+0.172, ' reckon':+0.162] -['Verify':-0.211, ' dzi':-0.190, ' Verify':-0.178]
L17U956 mean_up_read: +['aughan':+0.164, ' Vaughan':+0.163, 'homes':+0.158] -[' debunk':-0.179, ' acept':-0.172, ' boldly':-0.165]
L17U956 mean_output_write: +[' Jerusal':+0.148, ' Reject':+0.147, 'SEED':+0.143] -[' marin':-0.163, 'ALY':-0.137, ' afl':-0.135]
L17U956 B-A_output_write: +[' in':+0.290, ' un':+0.278, ' the':+0.275] -[' cinq':-0.256, ' SIX':-0.243, ' FOUR':-0.241]
L17U1149 mean_gate_read: +[' monstrous':+0.202, ' slayer':+0.184, ' devast':+0.179] -['ummies':-0.184, ' hyperfine':-0.166, 'documentElement':-0.158]
L17U1149 mean_up_read: +['ificacion':+0.191, ' aggiungere':+0.183, 'hingga':+0.177] -[' scholarships':-0.201, ' lawyers':-0.201, ' Lawyers':-0.193]
L17U1149 mean_output_write: +[' pemerintahan':+0.164, ' prosent':+0.140, ' alum':+0.132] -['getBy':-0.144, 'getView':-0.132, ' Breeding':-0.129]
L17U1149 B-A_output_write: +['LookAndFeelInfo':+0.267, 'LastGenOutput':+0.262, ' threx':+0.261] -[' the':-0.341, ' in':-0.330, ' one':-0.319]
