spb/wpai0001_uqo

1% source: experiments/summaries/graft_tier0_30e3644aa801/results.json (generated by scripts/build_tables.py)2\begin{tabular}{lccccc}3\toprule4 & \multicolumn{2}{c}{same initialization ($n=3$ pairs)} & \multicolumn{2}{c}{different initialization ($n=6$)} & control-adjusted \\5\cmidrule(lr){2-3} \cmidrule(lr){4-5}6Condition & target (sd) & control & target (sd) & control & $\Delta_T^{\mathrm{adj}}$ [95\% CI] \\7\midrule8Recipient, unmodified & 0.075 (0.03) & 1.000 & & & \\9Activation graft, identity map & 0.873 (0.10) & 0.990 & 0.101 (0.01) & 0.101 & +0.260 [+0.234, +0.287] \\10Activation graft, Procrustes (task rows) & 0.785 (0.08) & 0.985 & 0.697 (0.15) & 0.943 & +0.628 [+0.596, +0.660] \\11Activation graft, Procrustes (task-free rows) & 0.993 (0.01) & 0.996 & 0.956 (0.04) & 0.989 & +0.870 [+0.849, +0.891] \\12Activation graft, ridge (task-free rows) & 0.814 (0.08) & 1.000 & 0.795 (0.13) & 1.000 & +0.703 [+0.675, +0.732] \\13Activation graft, ridge (task rows) & 0.170 (0.08) & 0.994 & 0.175 (0.08) & 0.991 & +0.075 [+0.046, +0.104] \\14Activation graft, CCA (task rows) & 0.080 (0.03) & 1.000 & 0.089 (0.04) & 1.000 & -0.012 [-0.038, +0.013] \\15Control: sham map (random orthogonal) & 0.102 (0.01) & 0.085 & 0.096 (0.00) & 0.100 & +0.000 [+0.000, +0.000] \\16Control: norm-matched random rows & 0.089 (0.00) & 0.097 & 0.089 (0.00) & 0.099 & -0.009 [-0.034, +0.017] \\17Control: shuffled-item donor rows & 0.102 (0.00) & 0.101 & 0.099 (0.01) & 0.101 & +0.002 [-0.024, +0.028] \\18Control: donor layer 3 rows (ridge) & 0.187 (0.07) & 1.000 & 0.194 (0.08) & 1.000 & +0.093 [+0.064, +0.123] \\19Control: unrelated-skill donor (ridge) & 0.076 (0.04) & 1.000 & 0.076 (0.04) & 1.000 & -0.022 [-0.046, +0.003] \\20Steering vector, $\alpha=1$ & 0.064 (0.02) & 1.000 & ---  & --- & -0.037 [-0.062, -0.013] \\21Steering vector, $\alpha=4$ & 0.096 (0.00) & 0.611 & ---  & --- & -0.006 [-0.033, +0.019] \\22Subspace graft, $k=8$ & 0.936 (0.07) & 1.000 & ---  & --- & +0.834 [+0.811, +0.856] \\23Subspace graft, $k=32$ & 0.983 (0.02) & 1.000 & ---  & --- & +0.881 [+0.861, +0.901] \\24Subspace graft, $k=128$ & 0.993 (0.01) & 0.996 & ---  & --- & +0.891 [+0.872, +0.910] \\25Control: random subspace, $k=32$ & 0.173 (0.04) & 1.000 & ---  & --- & +0.071 [+0.043, +0.099] \\26Adapter graft (donor-free, ridge interface) & 0.581 (0.11) & 0.998 & ---  & --- & +0.480 [+0.445, +0.515] \\27Baseline: supervised adapter (labels) & 1.000 (0.00) & 1.000 & ---  & --- & +0.898 [+0.879, +0.916] \\28Baseline: distillation (donor outputs) & 1.000 (0.00) & 1.000 & ---  & --- & +0.898 [+0.879, +0.916] \\29Weight graft: L0 MLP & 0.262 (0.14) & 0.962 & ---  & --- & +0.160 [+0.129, +0.192] \\30Weight graft: L0 attention & 0.079 (0.03) & 0.733 & ---  & --- & -0.023 [-0.048, +0.003] \\31Weight graft: L0 block & 0.642 (0.19) & 0.726 & ---  & --- & +0.540 [+0.507, +0.574] \\32Weight graft: L0--L2 blocks + $\Omega$ embedding & 0.774 (0.26) & 0.749 & ---  & --- & +0.672 [+0.644, +0.699] \\33Weight graft: $\Omega$ embedding only & 0.075 (0.02) & 1.000 & ---  & --- & -0.027 [-0.051, -0.002] \\34Control: donor L3 block into L0 & 0.083 (0.01) & 0.089 & ---  & --- & -0.018 [-0.044, +0.007] \\35Control: unrelated-skill donor L0 block & 0.094 (0.04) & 0.933 & ---  & --- & -0.007 [-0.033, +0.019] \\36\bottomrule37\end{tabular}38