OPPOSING-HEAD PAIR STRUCTURE
The common/differential energy ratio compares (head_i+head_j)/2
with (head_i-head_j)/2 as WEIGHT maps, not numerical head outputs.
Metrics: raw O; residual OV; balanced OV; vocabulary-weighted OV.

L11 H2/H3
 A raw_O              : common/differential=0.037104, participation_rank=191.69, first5=[0.0214 0.0147 0.0086 0.0083 0.0082]
 A residual_OV        : common/differential=0.012148, participation_rank=31.77, first5=[0.0799 0.0668 0.0564 0.0537 0.0429]
 A balanced_OV        : common/differential=0.007225, participation_rank=72.29, first5=[0.032  0.0315 0.0299 0.0281 0.025 ]
 A vocab_weighted_OV  : common/differential=0.014322, participation_rank=30.75, first5=[0.0908 0.0747 0.0527 0.0451 0.039 ]
 B raw_O              : common/differential=0.037109, participation_rank=191.69, first5=[0.0214 0.0147 0.0086 0.0083 0.0082]
 B residual_OV        : common/differential=0.012162, participation_rank=31.76, first5=[0.0799 0.0669 0.0563 0.0537 0.0429]
 B balanced_OV        : common/differential=0.007232, participation_rank=72.29, first5=[0.032  0.0315 0.0299 0.0281 0.025 ]
 B vocab_weighted_OV  : common/differential=0.014373, participation_rank=30.76, first5=[0.0907 0.0746 0.0527 0.0452 0.039 ]
 query-matrix cosine A/B: 0.397331907263457 0.39747668686822046
 relative B-A differential-query matrix norm: 7.712841033935547e-05
 cosine between the two heads' query changes: -0.5875236781780236

L17 H0/H2
 A raw_O              : common/differential=0.273434, participation_rank=140.87, first5=[0.0304 0.0161 0.0125 0.0122 0.0116]
 A residual_OV        : common/differential=0.180842, participation_rank=32.07, first5=[0.0727 0.062  0.056  0.0524 0.047 ]
 A balanced_OV        : common/differential=0.159039, participation_rank=72.15, first5=[0.0375 0.0322 0.029  0.0268 0.026 ]
 A vocab_weighted_OV  : common/differential=0.185115, participation_rank=26.87, first5=[0.0964 0.0739 0.0553 0.0512 0.0479]
 B raw_O              : common/differential=0.273444, participation_rank=140.87, first5=[0.0304 0.0161 0.0125 0.0122 0.0116]
 B residual_OV        : common/differential=0.180846, participation_rank=32.07, first5=[0.0727 0.0619 0.056  0.0525 0.047 ]
 B balanced_OV        : common/differential=0.159070, participation_rank=72.14, first5=[0.0375 0.0322 0.029  0.0268 0.026 ]
 B vocab_weighted_OV  : common/differential=0.185398, participation_rank=26.86, first5=[0.0965 0.0739 0.0553 0.0511 0.0479]
 query-matrix cosine A/B: 0.549278640240232 0.548163045431084
 relative B-A differential-query matrix norm: 0.0015639066696166992
 cosine between the two heads' query changes: -0.667243690706066

L7 H0/H1
 A raw_O              : common/differential=0.295258, participation_rank=150.03, first5=[0.0323 0.0141 0.0134 0.0114 0.0112]
 A residual_OV        : common/differential=0.308047, participation_rank=2.72, first5=[0.5955 0.1116 0.0106 0.0084 0.0077]
 A balanced_OV        : common/differential=0.266169, participation_rank=19.04, first5=[0.2152 0.0244 0.0217 0.0201 0.0169]
 A vocab_weighted_OV  : common/differential=0.232445, participation_rank=6.94, first5=[0.3624 0.0561 0.0364 0.0338 0.0296]
 B raw_O              : common/differential=0.295245, participation_rank=150.03, first5=[0.0323 0.0141 0.0134 0.0114 0.0112]
 B residual_OV        : common/differential=0.308330, participation_rank=2.72, first5=[0.5948 0.1119 0.0106 0.0084 0.0077]
 B balanced_OV        : common/differential=0.266051, participation_rank=19.02, first5=[0.2154 0.0244 0.0216 0.0201 0.0169]
 B vocab_weighted_OV  : common/differential=0.232480, participation_rank=6.95, first5=[0.3619 0.0561 0.0365 0.0338 0.0296]
 query-matrix cosine A/B: 0.49357643767278264 0.49352101473086213
 relative B-A differential-query matrix norm: 7.259845733642578e-05
 cosine between the two heads' query changes: -0.3148081390153347

VOCABULARY-FACING FEATURES OF THE DIFFERENTIAL WRITING MAP
A is an arbitrary structural reference here, not an assumed starting model.

L11 H2/H3
 feature1 input-reading +: [("' suporte'", 4.003), ("' pasando'", 3.835), ("' gunung'", 3.801), ("' connaissances'", 3.777), ("' tumbuhan'", 3.772)]
 input-reading -: [("' PLAY'", -3.599), ("' Play'", -3.57), ("' Min'", -3.455), ("' play'", -3.407), ("' plays'", -3.368)]
 output-writing +: [("'alogical'", 3.89), ("'quoia'", 3.843), ("' poziomie'", 3.795), ("' abajo'", 3.76), ("' duvar'", 3.636)]
 output-writing -: [("' Chand'", -4.567), ("' Cricket'", -4.565), ("' Am'", -4.214), ("'Am'", -4.083), ("' FDI'", -3.983)]
 largest output coordinates: [481, 295, 637, 315] squared-coordinate share: 0.1834138035774231
 feature2 input-reading +: [("' uda'", 4.354), ("'sohn'", 4.183), ("' inverso'", 4.169), ("' wajah'", 4.099), ("' escenas'", 4.009)]
 input-reading -: [("' inflow'", -4.048), ("' magnes'", -3.941), ("' pdb'", -3.874), ("' restricting'", -3.814), ("'ahraga'", -3.785)]
 output-writing +: [("' test'", 4.461), ("' oil'", 4.416), ("' tester'", 4.215), ("' Emma'", 4.171), ("' Test'", 4.069)]
 output-writing -: [("' maddenin'", -4.04), ("' Villar'", -3.877), ("' Whiting'", -3.834), ("' Vitor'", -3.751), ("' Burnu'", -3.751)]
 largest output coordinates: [490, 472, 315, 16] squared-coordinate share: 0.1840747445821762

L17 H0/H2
 feature1 input-reading +: [("' radicales'", 4.081), ("' afectado'", 4.024), ("' verdadera'", 3.782), ("' semillas'", 3.755), ("' citado'", 3.75)]
 input-reading -: [("' seabed'", -3.442), ("' plays'", -3.42), ("' submarine'", -3.323), ("' shipyard'", -3.295), ("' sonar'", -3.287)]
 output-writing +: [("' rurales'", 4.918), ("' montagna'", 4.737), ("' antlers'", 4.613), ("' Pronghorn'", 4.529), ("'deer'", 4.504)]
 output-writing -: [("' Caribbean'", -4.488), ("'Roberto'", -4.429), ("'Caribbean'", -4.333), ("' Rizzo'", -4.237), ("' Crist'", -4.215)]
 largest output coordinates: [517, 327, 163, 320] squared-coordinate share: 0.1832108050584793

L7 H0/H1
 feature1 input-reading +: [("' avanzar'", 3.474), ("' enemigo'", 3.304), ("' directorio'", 3.269), ("'leyici'", 3.226), ("' apoyo'", 3.222)]
 input-reading -: [("' LED'", -2.707), ("' Max'", -2.706), ("' EU'", -2.68), ("' Johnny'", -2.63), ("' Meg'", -2.608)]
 output-writing +: [("' cuad'", 3.32), ("' directorio'", 3.185), ("' designado'", 3.177), ("' grueso'", 3.16), ("'aksanakan'", 3.092)]
 output-writing -: [("' LED'", -3.019), ("' EU'", -2.855), ("' Smart'", -2.649), ("' Sp'", -2.644), ("' Meg'", -2.608)]
 largest output coordinates: [163, 400, 87, 188] squared-coordinate share: 0.7966848611831665

L11H2/H3: CHANGE IN THE DIFFERENTIAL QUERY/KEY OPERATOR
Separate contribution squared norms / total squared norm: 0.646603630817531 0.36511916765244595
Their sum need not be one because the contributions have a cross term.
Leading differential-change energy fractions: [0.3827 0.134  0.0936 0.058 ]
 change-factor1 output-query +: [("' Abu'", 3.918), ("'lining'", 3.912), ("' Tank'", 3.891), ("'Abu'", 3.854), ("' Casablanca'", 3.847)] -: [("'9'", -5.507), ("'8'", -4.35), ("'7'", -4.298), ("'5'", -4.249), ("'4'", -4.116)]
 source-key +: [("' inhomogeneities'", 4.537), ("'ouble'", 4.001), ("' matchmaking'", 3.951), ("' Fonbet'", 3.867), ("'ulator'", 3.838)] -: [("'5'", -9.553), ("'7'", -9.478), ("'9'", -9.106), ("'6'", -8.959), ("'4'", -8.774)]
 change-factor2 output-query +: [("' Abre'", 4.056), ("' Painlev'", 3.962), ("' accomod'", 3.892), ("' Accom'", 3.89), ("' Happiness'", 3.809)] -: [("'3'", -6.965), ("'7'", -6.661), ("'5'", -6.531), ("'4'", -6.499), ("'8'", -6.193)]
 source-key +: [("'quet'", 4.687), ("' bescherm'", 4.242), ("'purs'", 4.163), ("' binatang'", 4.139), ("'oud'", 4.016)] -: [("' Installed'", -4.461), ("' Comptroller'", -4.412), ("' instrumentality'", -4.104), ("' electrician'", -4.021), ("' valorem'", -3.91)]

No attention pattern, retrieved mixture, or head activation was computed.
All comparisons concern weight operators and vocabulary-weight projections.
