Target heads: [(11, 2), (10, 2)] ; fixed relative distances: [9, 24]
MLPs 0..9 precede both targets; MLP 10 precedes L11 but not L10.
The main comparable ranking therefore uses MLPs 0..9 only.
Only changed down-projection/output-normalization weights are screened here.
Unknown MLP activations, gates, RMS denominators and intervening transformations are omitted.
Enrichment = projected change energy / expected energy after independent residual-coordinate sign flips.

L  raw_key(relative,enrichment) raw_query(relative,enrichment) scaled_key(relative,enrichment) scaled_query(relative,enrichment)
 0 (0.01057,1.000) (0.01069,1.026) (0.01353,1.025) (0.01402,1.135)
 1 (0.01653,1.032) (0.01661,1.046) (0.01993,1.083) (0.01939,1.052)
 2 (0.01392,1.050) (0.01386,1.048) (0.01513,1.057) (0.01487,1.034)
 3 (0.01613,1.059) (0.01556,1.005) (0.01852,1.150) (0.01816,1.101)
 4 (0.01849,1.071) (0.01850,1.074) (0.02302,1.173) (0.02264,1.114)
 5 (0.01104,1.119) (0.01080,1.110) (0.01172,1.178) (0.01156,1.192)
 6 (0.01517,1.142) (0.01463,1.091) (0.01586,1.195) (0.01485,1.132)
 7 (0.01413,1.164) (0.01308,1.147) (0.01450,1.165) (0.01340,1.190)
 8 (0.01999,1.323) (0.01805,1.188) (0.02090,1.305) (0.01925,1.231)
 9 (0.01731,1.437) (0.01458,1.493) (0.01924,1.453) (0.01636,1.620)

RANKING SENSITIVITY
raw key relative [(8, np.float64(0.01999)), (4, np.float64(0.01849)), (9, np.float64(0.01731)), (1, np.float64(0.01653)), (3, np.float64(0.01613))]
raw key enrichment [(9, 1.43742), (8, 1.32267), (7, 1.16383), (6, 1.14221), (5, 1.11909)]
raw key projected_norm [(9, np.float64(0.01231)), (8, np.float64(0.01127)), (5, np.float64(0.01089)), (6, np.float64(0.01079)), (7, np.float64(0.01075))]
raw query relative [(4, np.float64(0.0185)), (8, np.float64(0.01805)), (1, np.float64(0.01661)), (3, np.float64(0.01556)), (6, np.float64(0.01463))]
raw query enrichment [(9, 1.49275), (8, 1.18762), (7, 1.14729), (5, 1.11048), (6, 1.09104)]
raw query projected_norm [(9, np.float64(0.01238)), (5, np.float64(0.01082)), (8, np.float64(0.01064)), (7, np.float64(0.01064)), (4, np.float64(0.01054))]
scaled key relative [(4, np.float64(0.02302)), (8, np.float64(0.0209)), (1, np.float64(0.01993)), (9, np.float64(0.01924)), (3, np.float64(0.01852))]
scaled key enrichment [(9, 1.45256), (8, 1.30471), (6, 1.19524), (5, 1.17799), (4, 1.17251)]
scaled key projected_norm [(9, np.float64(0.18749)), (8, np.float64(0.10841)), (7, np.float64(0.08877)), (6, np.float64(0.07349)), (5, np.float64(0.05744))]
scaled query relative [(4, np.float64(0.02264)), (1, np.float64(0.01939)), (8, np.float64(0.01925)), (3, np.float64(0.01816)), (9, np.float64(0.01636))]
scaled query enrichment [(9, 1.61956), (8, 1.23099), (5, 1.19188), (7, 1.18957), (0, 1.13454)]
scaled query projected_norm [(9, np.float64(0.16429)), (8, np.float64(0.08948)), (7, np.float64(0.07752)), (6, np.float64(0.0632)), (3, np.float64(0.05955))]

UNIT CANDIDATES FROM ROUTING-SUBSPACE PROJECTIONS
Scores describe changed columns, not measured unit activity or behavioral effect.
MLP 1
  raw key [(802, 0.00111), (871, 0.0011), (1954, 0.00105), (306, 0.00101), (592, 0.00098)] top5_energy_fraction 0.00524
  raw query [(2023, 0.00114), (802, 0.00106), (1165, 0.00103), (378, 0.00102), (564, 0.00101)] top5_energy_fraction 0.00525
  scaled key [(1406, 0.00152), (1355, 0.00133), (425, 0.00131), (1640, 0.00128), (698, 0.00126)] top5_energy_fraction 0.0067
  scaled query [(1406, 0.00232), (698, 0.00157), (1901, 0.00152), (1640, 0.00137), (1285, 0.00137)] top5_energy_fraction 0.00815
MLP 4
  raw key [(1864, 0.00113), (595, 0.00111), (1959, 0.00104), (1162, 0.00104), (1908, 0.00103)] top5_energy_fraction 0.00536
  raw query [(595, 0.00119), (147, 0.0011), (1046, 0.00106), (816, 0.00104), (1669, 0.00103)] top5_energy_fraction 0.00543
  scaled key [(1336, 0.00141), (1021, 0.00122), (1949, 0.00121), (1538, 0.00117), (232, 0.00117)] top5_energy_fraction 0.00618
  scaled query [(595, 0.0013), (578, 0.0012), (968, 0.00116), (1427, 0.00116), (341, 0.00111)] top5_energy_fraction 0.00592
MLP 8
  raw key [(1098, 0.0013), (1546, 0.00119), (949, 0.00114), (1831, 0.00112), (1911, 0.00111)] top5_energy_fraction 0.00586
  raw query [(1030, 0.00137), (1466, 0.00121), (464, 0.00109), (446, 0.00104), (1521, 0.00103)] top5_energy_fraction 0.00574
  scaled key [(909, 0.00144), (1098, 0.00121), (1751, 0.00116), (1466, 0.00109), (1571, 0.00108)] top5_energy_fraction 0.00599
  scaled query [(1571, 0.00121), (1751, 0.00114), (1236, 0.00111), (297, 0.00111), (711, 0.00103)] top5_energy_fraction 0.0056

An aligned weight change may be inactive, canceled, or transformed downstream.
