\begin{table}[t]
\centering\scriptsize
\caption{Statistical audit: medians with bootstrap 95\,\% CIs (10\,000 resamples) of decode throughput (tok/s) and of the difference A $-$ B.}
\label{tab:audit}
\begin{tabular}{L{7.2cm}rrrr}
\toprule
Comparison (A vs B) & $n_A/n_B$ & A [CI] & B [CI] & A$-$B [CI] \\
\midrule
laptop: TITAN all-resident (64 slots, preloaded) vs mlx-lm & 12/12 & 489.1 [488.2, 489.5] & 495.0 [494.1, 495.9] & -5.9 [-7.1, -4.8] \\
laptop: TITAN 64 slots not preloaded (sync path) vs mlx-lm & 12/12 & 201.4 [201.0, 202.7] & 495.0 [494.1, 495.9] & -293.6 [-294.4, -292.0] \\
M3U96a Qwen3-30B: TITAN all-resident (128 slots) vs mlx-lm & 18/18 & 121.6 [121.4, 121.8] & 121.7 [121.4, 121.8] & -0.1 [-0.3, +0.3] \\
laptop cold NVMe, 16 slots: transition m=8 vs no prefetch & 12/12 & 27.9 [26.5, 30.5] & 36.4 [34.1, 42.0] & -8.5 [-14.5, -5.1] \\
laptop cold NVMe, 16 slots: transition2 m=8 vs no prefetch & 12/12 & 35.0 [33.7, 41.5] & 36.4 [34.1, 42.0] & -1.4 [-7.3, +6.0] \\
laptop cold NVMe, 16 slots: transition3 m=8 vs no prefetch & 12/12 & 29.1 [27.7, 34.7] & 36.4 [34.1, 42.0] & -7.3 [-13.1, -1.1] \\
laptop cold NVMe, 16 slots: transition2 m=16 vs no prefetch & 12/12 & 25.9 [25.0, 29.6] & 36.4 [34.1, 42.0] & -10.4 [-16.1, -5.9] \\
laptop cold NVMe, 16 slots: prev\_rank m=16 vs no prefetch & 12/12 & 19.1 [18.7, 20.0] & 36.4 [34.1, 42.0] & -17.2 [-22.9, -14.8] \\
laptop cold NVMe, 32 slots: transition m=8 vs no prefetch & 12/12 & 41.3 [35.5, 46.9] & 50.6 [46.6, 56.9] & -9.2 [-17.5, -2.6] \\
laptop cold NVMe, 32 slots: transition2 m=8 vs no prefetch & 12/12 & 42.0 [38.5, 46.8] & 50.6 [46.6, 56.9] & -8.5 [-15.2, -2.7] \\
laptop cold NVMe, 32 slots: transition3 m=8 vs no prefetch & 12/12 & 41.0 [37.6, 46.6] & 50.6 [46.6, 56.9] & -9.5 [-16.2, -3.0] \\
laptop cold NVMe, 32 slots: transition2 m=16 vs no prefetch & 12/12 & 41.0 [35.4, 46.0] & 50.6 [46.6, 56.9] & -9.6 [-17.7, -3.6] \\
laptop cold NVMe, 32 slots: prev\_rank m=16 vs no prefetch & 12/12 & 28.4 [27.0, 32.5] & 50.6 [46.6, 56.9] & -22.2 [-28.6, -16.9] \\
laptop cold NVMe, 16 slots (general): probe m=8 vs no prefetch & 4/4 & 31.8 [31.6, 32.8] & 33.0 [32.6, 33.3] & -1.2 [-1.6, -0.2] \\
laptop cold NVMe, 16 slots (general): probe m=16 vs no prefetch & 4/4 & 25.3 [24.8, 26.2] & 33.0 [32.6, 33.3] & -7.7 [-8.4, -6.8] \\
laptop cold NVMe, 16 slots (general): probe2 m=8 vs no prefetch & 4/4 & 31.3 [31.0, 32.0] & 33.0 [32.6, 33.3] & -1.7 [-2.1, -0.9] \\
laptop cold NVMe, 16 slots (general): probe3 m=8 vs no prefetch & 4/4 & 30.7 [29.3, 32.6] & 33.0 [32.6, 33.3] & -2.3 [-3.7, -0.4] \\
laptop cold NVMe, 16 slots (general): transition2 m=8 vs no prefetch & 4/4 & 27.1 [26.0, 28.2] & 33.0 [32.6, 33.3] & -5.9 [-7.1, -4.5] \\
laptop cold NVMe, 16 slots (general): transition2t50 m=8 vs no prefetch & 4/4 & 27.9 [26.8, 29.4] & 33.0 [32.6, 33.3] & -5.1 [-6.2, -3.6] \\
laptop cold NVMe, 16 slots (general): transition2t70 m=8 vs no prefetch & 4/4 & 27.8 [26.7, 29.4] & 33.0 [32.6, 33.3] & -5.2 [-6.3, -3.6] \\
laptop H4: fused up+gate vs two kernels, 64 slots resident & 12/12 & 475.6 [473.5, 476.2] & 476.7 [475.7, 478.2] & -1.1 [-3.6, +0.1] \\
laptop H4: fused up+gate vs two kernels, 16 slots cold & 4/4 & 35.9 [35.7, 36.7] & 35.0 [34.1, 35.5] & +1.0 [+0.4, +2.0] \\
laptop cold NVMe, 16 slots, 256-token generations (energy run): probe m=8 vs no prefetch & 12/12 & 42.0 [39.1, 45.1] & 31.9 [26.8, 35.2] & +10.1 [+5.2, +16.9] \\
laptop energy run: TITAN all-resident vs mlx-lm (256 tokens) & 12/12 & 461.9 [460.0, 463.3] & 463.0 [457.2, 467.4] & -1.1 [-6.3, +4.9] \\
M3U96a+M3U96b OLMoE 32 slots: odd experts remote vs local & 16/16 & 15.1 [14.8, 15.5] & 81.3 [70.5, 86.9] & -66.2 [-71.8, -55.2] \\
M3U96a+M3U96b Qwen3-30B 32 slots cold: odd placement vs local cold NVMe & 16/16 & 4.2 [4.2, 4.3] & 8.8 [8.3, 9.3] & -4.5 [-5.1, -4.0] \\
M3U96a+M3U96b Qwen3-30B 32 slots cold: popular placement vs local cold NVMe & 16/16 & 5.1 [4.6, 5.6] & 8.8 [8.3, 9.3] & -3.6 [-4.4, -2.9] \\
M3U96a+M3U96b Qwen3-30B 64 slots cold: odd placement vs local cold NVMe & 16/16 & 5.2 [5.2, 5.3] & 16.3 [14.9, 19.7] & -11.1 [-14.5, -9.7] \\
M3U96a+M3U96b Qwen3-30B 64 slots cold: popular placement vs local cold NVMe & 16/16 & 6.0 [5.8, 6.7] & 16.3 [14.9, 19.7] & -10.2 [-13.6, -8.8] \\
\bottomrule
\end{tabular}
\end{table}