spb/wpai0001_uqo

1% source: experiments/summaries/graft_tier0_confirmatory_0c2257f6803e/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 & 0.071 (0.02) & 1.000 & & & \\9Activation graft, identity map & 0.868 (0.10) & 0.989 & 0.095 (0.01) & 0.102 & +0.262 [+0.235, +0.287] \\10Activation graft, Procrustes (task rows) & 0.790 (0.10) & 0.983 & 0.697 (0.16) & 0.936 & +0.636 [+0.605, +0.667] \\11Activation graft, Procrustes (task-free rows) & 0.995 (0.01) & 0.999 & 0.952 (0.04) & 0.988 & +0.875 [+0.854, +0.895] \\12Activation graft, ridge (task-free rows) & 0.802 (0.10) & 1.000 & 0.788 (0.13) & 1.000 & +0.702 [+0.672, +0.731] \\13Activation graft, ridge (task rows) & 0.173 (0.06) & 0.994 & 0.175 (0.08) & 0.988 & +0.083 [+0.054, +0.111] \\14Activation graft, CCA (task rows) & 0.076 (0.03) & 0.999 & 0.083 (0.04) & 1.000 & -0.010 [-0.035, +0.014] \\15Control: sham map (random orthogonal) & 0.089 (0.00) & 0.100 & 0.093 (0.01) & 0.100 & +0.000 [+0.000, +0.000] \\16Control: norm-matched random rows & 0.091 (0.01) & 0.093 & 0.091 (0.01) & 0.093 & -0.000 [-0.026, +0.025] \\17Control: shuffled-item donor rows & 0.106 (0.00) & 0.098 & 0.101 (0.00) & 0.097 & +0.011 [-0.014, +0.037] \\18Control: donor layer 3 rows (ridge) & 0.186 (0.08) & 1.000 & 0.187 (0.09) & 1.000 & +0.095 [+0.066, +0.124] \\19Control: unrelated-skill donor (ridge) & 0.076 (0.03) & 1.000 & 0.076 (0.03) & 1.000 & -0.015 [-0.039, +0.009] \\20Steering vector, $\alpha=1$ & 0.066 (0.02) & 1.000 & ---  & --- & -0.023 [-0.046, +0.001] \\21Steering vector, $\alpha=4$ & 0.090 (0.01) & 0.598 & ---  & --- & +0.002 [-0.024, +0.027] \\22Subspace graft, $k=8$ & 0.927 (0.09) & 1.000 & ---  & --- & +0.839 [+0.815, +0.861] \\23Subspace graft, $k=32$ & 0.983 (0.02) & 1.000 & ---  & --- & +0.894 [+0.875, +0.913] \\24Subspace graft, $k=128$ & 0.995 (0.01) & 0.999 & ---  & --- & +0.906 [+0.887, +0.924] \\25Control: random subspace, $k=32$ & 0.168 (0.05) & 1.000 & ---  & --- & +0.079 [+0.051, +0.107] \\26Adapter graft (donor-free, ridge interface) & 0.544 (0.13) & 0.998 & ---  & --- & +0.455 [+0.422, +0.490] \\27Baseline: supervised adapter (labels) & 1.000 (0.00) & 1.000 & ---  & --- & +0.911 [+0.892, +0.928] \\28Baseline: distillation (donor outputs) & 1.000 (0.00) & 1.000 & ---  & --- & +0.911 [+0.893, +0.929] \\29Weight graft: L0 MLP & 0.245 (0.13) & 0.959 & ---  & --- & +0.156 [+0.127, +0.187] \\30Weight graft: L0 attention & 0.087 (0.02) & 0.732 & ---  & --- & -0.002 [-0.027, +0.022] \\31Weight graft: L0 block & 0.642 (0.19) & 0.720 & ---  & --- & +0.553 [+0.520, +0.586] \\32Weight graft: L0--L2 blocks + $\Omega$ embedding & 0.778 (0.26) & 0.751 & ---  & --- & +0.689 [+0.662, +0.715] \\33Weight graft: $\Omega$ embedding only & 0.073 (0.01) & 1.000 & ---  & --- & -0.016 [-0.039, +0.008] \\34Control: donor L3 block into L0 & 0.086 (0.01) & 0.086 & ---  & --- & -0.003 [-0.027, +0.022] \\35Control: unrelated-skill donor L0 block & 0.097 (0.04) & 0.931 & ---  & --- & +0.008 [-0.018, +0.033] \\36\bottomrule37\end{tabular}38