OUTPUT of run 20 of at most 60:
```
COMMON INPUT-FEATURE REMAPPING TEST
For stacked input projections, M=(A+B)/2 and D=B-A.
Decomposition: D = M@T + residual, using an algebraic pseudoinverse.
The state-dependent normalization factors are NOT evaluated.

L0_attention: retained rank=640, Gram condition=1848
 projected difference energy=0.50612; entry-sign-null expectation=0.51182; orthogonal residual=0.49388
 projection/residual inner product divided by energy=-3.47e-16

L0_mlp: retained rank=639, Gram condition=834.9
 projected difference energy=0.15612; entry-sign-null expectation=0.15672; orthogonal residual=0.84388
 projection/residual inner product divided by energy=-3.15e-17

L1_attention: retained rank=640, Gram condition=3.542e+05
 projected difference energy=0.56730; entry-sign-null expectation=0.57013; orthogonal residual=0.43270
 projection/residual inner product divided by energy=+1.33e-15

L1_mlp: retained rank=638, Gram condition=3600
 projected difference energy=0.15502; entry-sign-null expectation=0.15651; orthogonal residual=0.84498
 projection/residual inner product divided by energy=-7.17e-17

L2_mlp: retained rank=639, Gram condition=1.689e+04
 projected difference energy=0.15686; entry-sign-null expectation=0.15695; orthogonal residual=0.84314
 projection/residual inner product divided by energy=+3.26e-16

CROSS-BRANCH ALGEBRAIC CONSISTENCY
Score = 1 - ||D_target - M_target@T_source||^2 / ||D_target||^2.
This compares weight matrices only. Negative scores mean the shared-map
interpretation worsens the target-branch residual relative to no remapping.
L0_mlp: gate->up=-0.51875, up->gate=-0.54867, map cosine=-0.00616
 gate-only projected energy/null=0.31610/0.31353; up-only=0.30933/0.31313
L1_mlp: gate->up=-0.51213, up->gate=-0.50014, map cosine=-0.02140
 gate-only projected energy/null=0.31380/0.31291; up-only=0.31528/0.31255
L2_mlp: gate->up=-0.49890, up->gate=-0.45934, map cosine=+0.03542
 gate-only projected energy/null=0.31770/0.31410; up-only=0.31060/0.31295

SEMANTIC DIRECTIONS OF THE PROJECTED INPUT REMAPPING
T @ embedding-row is a weight-derived displacement for the explainable
part only. It is NOT an actual embedding update or a hidden-state change.

 L0_attention
 ' think': displacement norm=0.044553
  +: [("' Pierre'", 0.2344), ("' includes'", 0.2257), ("' depicts'", 0.2192), ("' stipulates'", 0.2145), ("'findOne'", 0.2143)]
  -: [("' nenhum'", -0.1436), ("' unaffected'", -0.129), ("' all'", -0.1276), ("'ukiy'", -0.1241), ("'quia'", -0.1218)]
 ' answer': displacement norm=0.055220
  +: [("' raster'", 0.2025), ("'Robert'", 0.202), ("'Jan'", 0.2013), ("'Jesus'", 0.198), ("'Marc'", 0.1979)]
  -: [("' nenhum'", -0.1751), ("' INTERRUPTION'", -0.1464), ("'ukiy'", -0.1384), ("'banding'", -0.1364), ("' sark'", -0.1344)]
 ' banana': displacement norm=0.072229
  +: [("' presents'", 0.1776), ("'riminal'", 0.1649), ("' Verantwort'", 0.1639), ("' includes'", 0.1628), ("' respons'", 0.1619)]
  -: [("' ditambah'", -0.149), ("' exclusivo'", -0.1486), ("' nenhum'", -0.1448), ("'quia'", -0.1409), ("' all'", -0.1407)]

 L0_mlp
 ' think': displacement norm=0.013584
  +: [("' Children'", 0.1885), ("' wins'", 0.1811), ("' Child'", 0.1809), ("' Fear'", 0.1769), ("' Life'", 0.1746)]
  -: [("'erjee'", -0.1355), ("'cropping'", -0.1287), ("'liess'", -0.1207), ("' foreclose'", -0.1186), ("'slu'", -0.1146)]
 ' answer': displacement norm=0.017031
  +: [("' ethernet'", 0.1977), ("' winners'", 0.1849), ("' Up'", 0.1847), ("' EB'", 0.1802), ("' Fin'", 0.1794)]
  -: [("'voiced'", -0.1336), ("' consultar'", -0.1258), ("'chloric'", -0.1218), ("' desenvol'", -0.1216), ("' ortaya'", -0.1209)]
 ' banana': displacement norm=0.016876
  +: [("' health'", 0.1898), ("' sanity'", 0.1856), ("' metrics'", 0.1839), ("' quantitative'", 0.1709), ("' physical'", 0.1708)]
  -: [("' venne'", -0.1427), ("' patente'", -0.137), ("' terdapat'", -0.1346), ("' proposent'", -0.1317), ("'warn'", -0.1298)]

 L1_attention
 ' think': displacement norm=0.166784
  +: [("'setPrototypeOf'", 0.2213), ("'TestSuite'", 0.1867), ("'itriangular'", 0.1664), ("'acute'", 0.1641), ("'roweak'", 0.1617)]
  -: [("'beitung'", -0.2045), ("'nlm'", -0.1981), ("'labelledby'", -0.1946), ("'dni'", -0.1936), ("'dropna'", -0.1874)]
 ' answer': displacement norm=0.192091
  +: [("'setPrototypeOf'", 0.2108), ("'ccak'", 0.1817), ("'TestSuite'", 0.162), ("'StdHandle'", 0.1532), ("'emonic'", 0.1486)]
  -: [("'zkod'", -0.1841), ("'udni'", -0.1827), ("'diphenyl'", -0.1774), ("'GORITHM'", -0.1767), ("' podej'", -0.1754)]
 ' banana': displacement norm=0.103507
  +: [("'newblock'", 0.1955), ("'rophe'", 0.1926), ("'istently'", 0.1862), ("'iori'", 0.17), ("'elessly'", 0.1677)]
  -: [("'setPrototypeOf'", -0.202), ("'Om'", -0.1888), ("'StdHandle'", -0.1789), ("' MY'", -0.1783), ("' ce'", -0.1746)]

 L1_mlp
 ' think': displacement norm=0.023923
  +: [("' aree'", 0.1505), ("'naires'", 0.1394), ("' ahead'", 0.1373), ("' sopra'", 0.1353), ("'istika'", 0.1342)]
  -: [("'ndan'", -0.1996), ("'pid'", -0.1993), ("'px'", -0.1877), ("'gehen'", -0.1853), ("'tete'", -0.185)]
 ' answer': displacement norm=0.041510
  +: [("'i'", 0.2297), ("'y'", 0.2128), ("'a'", 0.2065), ("'is'", 0.1944), ("'o'", 0.1902)]
  -: [("'RETURNTRANSFER'", -0.2499), ("'debugElement'", -0.2458), ("' stargazerCount'", -0.2353), ("'HeightSizeMode'", -0.233), ("'BuildAction'", -0.2282)]
 ' banana': displacement norm=0.019290
  +: [("'er'", 0.2124), ("'o'", 0.2003), ("'ing'", 0.1914), ("'a'", 0.1862), ("'al'", 0.1782)]
  -: [("'ousedown'", -0.1947), ("'ngModel'", -0.1929), ("'Eksterne'", -0.1843), ("'streamlit'", -0.178), ("'ContainsKey'", -0.1775)]

 L2_mlp
 ' think': displacement norm=0.029218
  +: [("'arsipkan'", 0.1834), ("' tabella'", 0.1588), ("'arschijnlijk'", 0.1515), ("'DOCKED'", 0.1507), ("'WhiteElo'", 0.1473)]
  -: [("'dar'", -0.1674), ("'uir'", -0.1451), ("'They'", -0.1374), ("'them'", -0.1373), ("'mn'", -0.1371)]
 ' answer': displacement norm=0.023561
  +: [("' keltiradi'", 0.1513), ("'getStartState'", 0.1507), ("'advant'", 0.1423), ("'endela'", 0.1352), ("'ock'", 0.1347)]
  -: [("'isez'", -0.1418), ("' taea'", -0.1389), ("'iserum'", -0.1389), ("'enciaga'", -0.1379), ("'PLOAD'", -0.1349)]
 ' banana': displacement norm=0.019213
  +: [("'tsky'", 0.1955), ("'Prosec'", 0.1825), ("' fielded'", 0.1751), ("'varlak'", 0.1747), ("'pointA'", 0.1745)]
  -: [("'s'", -0.2229), ("'a'", -0.1292), ("'t'", -0.0929), ("'sse'", -0.092), ("'sites'", -0.0899)]

DOMINANT REMAPPING FACTORS FOR THE TWO FIRST-LAYER BLOCKS

 L0_attention first three energy fractions: [0.1749 0.0882 0.0561]
 factor 1 input +: [("' California'", 0.624), ("' Florida'", 0.6204), ("' Israel'", 0.6203), ("' Miami'", 0.6154)] -: [("'temuan'", -0.0312), ("'ungkinan'", -0.0003), ("'hitungan'", 0.0029), ("'entingan'", 0.0052)]
 factor 1 displacement +: [("' includes'", 0.1931), ("' presents'", 0.1895), ("' Richard'", 0.1762), ("' Jonathan'", 0.176)] -: [("' nenhum'", -0.1573), ("' unaffected'", -0.1556), ("'wpi'", -0.1485), ("' conm'", -0.142)]
 factor 2 input +: [("' is'", 0.5348), ("' in'", 0.4635), ("' and'", 0.4415), ("' the'", 0.4326)] -: [("' RtlOverlay'", -0.5803), ("' llrp'", -0.575), ("' HUOBI'", -0.5699), ("' BVHTree'", -0.5664)]
 factor 2 displacement +: [("'setInt'", 0.2437), ("'JComboBox'", 0.2238), ("' REQUIREMENTS'", 0.2233), ("'Gson'", 0.2223)] -: [("'at'", -0.1548), ("'demon'", -0.1506), ("'med'", -0.1453), ("' repr'", -0.1434)]

 L0_mlp first three energy fractions: [0.118  0.084  0.0488]
 factor 1 input +: [("' Uruguay'", 0.5984), ("' Bolivia'", 0.5953), ("' Argentina'", 0.5883), ("' Guatemala'", 0.5868)] -: [("'1'", -0.3013), ("'2'", -0.2658), ("'3'", -0.2501), ("'0'", -0.2139)]
 factor 1 displacement +: [("' sanity'", 0.1886), ("'setInt'", 0.1863), ("' endIndex'", 0.1759), ("' pecuniary'", 0.1729)] -: [("'worlds'", -0.1332), ("'again'", -0.1264), ("' venne'", -0.1258), ("'Ltd'", -0.1196)]
 factor 2 input +: [("' super'", 0.514), ("' color'", 0.5108), ("' B'", 0.4896), ("' re'", 0.4857)] -: [("' llrp'", -0.3221), ("' renderEncoder'", -0.3171), ("' Tjiwarl'", -0.3171), ("' ivvu'", -0.3139)]
 factor 2 displacement +: [("' years'", 0.1525), ("' ambos'", 0.1514), ("' pubblico'", 0.1406), ("' ambas'", 0.1371)] -: [("'extract'", -0.1708), ("'suggestions'", -0.1669), ("' distract'", -0.1652), ("'boomerang'", -0.1647)]

No model parameters were changed, no training was performed, and no
sequence, activation function, gate, hidden-state normalization, or model output was run.

```