RADIAL FRACTIONS OF RAW WEIGHT-DIFFERENCE ENERGY
Attention O is treated as a WHOLE block: independently rescaling one head
would generally not be canceled by post-attention normalization.
Columns: layer | Q-head fractions | shared K fraction | whole O | whole MLP down
L00 | [0.     0.     0.     0.0001] | 0.000000 0.000001 0.000000
L01 | [0. 0. 0. 0.] | 0.000007 0.000076 0.000001
L02 | [0. 0. 0. 0.] | 0.000003 0.000002 0.000002
L03 | [0.     0.     0.     0.0001] | 0.000004 0.000002 0.000005
L04 | [0. 0. 0. 0.] | 0.000010 0.000003 0.000000
L05 | [0.     0.     0.     0.0003] | 0.000000 0.000008 0.000000
L06 | [0. 0. 0. 0.] | 0.000010 0.000010 0.000000
L07 | [0. 0. 0. 0.] | 0.000004 0.000000 0.000000
L08 | [0. 0. 0. 0.] | 0.000007 0.000005 0.000000
L09 | [0.     0.     0.0001 0.    ] | 0.000015 0.000010 0.000000
L10 | [0. 0. 0. 0.] | 0.000002 0.000008 0.000000
L11 | [0. 0. 0. 0.] | 0.000022 0.000002 0.000001
L12 | [0.0004 0.0001 0.     0.    ] | 0.000006 0.000000 0.000002
L13 | [0. 0. 0. 0.] | 0.000003 0.000007 0.000002
L14 | [0. 0. 0. 0.] | 0.000033 0.000003 0.000000
L15 | [0. 0. 0. 0.] | 0.000002 0.000000 0.000002
L16 | [0. 0. 0. 0.] | 0.000002 0.000001 0.000001
L17 | [0.0001 0.     0.0001 0.    ] | 0.000034 0.000000 0.000001

ATTENTION CANDIDATES: ORIGINAL / RADIAL-REMOVED COUPLINGS
All numbers are B-A weight couplings multiplied by 10^4.
Each pair uses the SAME original matrix-norm denominator.
L00 H3
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [ 0.8901  0.6421 -0.0281  0.0728 -0.0135]
  residual: [ 0.8898  0.642  -0.0281  0.0728 -0.0134]
L01 H1
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [-0.4458 -0.1158  0.0054  0.0917  0.0426]
  residual: [-0.4459 -0.1158  0.0054  0.0915  0.0429]
L05 H0
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [0.3099 0.3482 0.0082 0.0274 0.0026]
  residual: [0.3099 0.3481 0.0082 0.0274 0.0026]
L06 H0
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [ 0.2215  0.2442  0.0418 -0.0014 -0.0069]
  residual: [ 0.2215  0.2437  0.0421 -0.0011 -0.0066]
L16 H3
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [ 0.1325  0.1449  0.0285  0.0595 -0.0519]
  residual: [ 0.1324  0.1447  0.0286  0.0596 -0.0519]
L17 H2
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [-0.084   0.06    0.0827  0.1117  0.037 ]
  residual: [-0.0831  0.0595  0.0827  0.1117  0.0369]
L17 H3
  names: belief-question, belief-fact, ABCD, digits_1_9, boolean
  original: [-0.1466 -0.0204 -0.0418 -0.0465  0.0092]
  residual: [-0.1465 -0.0204 -0.0418 -0.0465  0.0092]

TERMINAL MLP UNGATED CONNECTIONS: ORIGINAL / RADIAL-REMOVED
L16: categories=['ABCD', 'digits_1_9', 'boolean']
  original: [ 0.0434 -0.2082  0.0493]
  residual: [ 0.0434 -0.2084  0.0492]
L17: categories=['ABCD', 'digits_1_9', 'boolean']
  original: [ 0.0855  0.5146 -0.1353]
  residual: [ 0.0853  0.5145 -0.1354]

This separates parameter directions only; no original or altered model was run.
