Table 5.

Paired statistical comparison of CarLab 2025 against each reference taxonomy. Inputs are taken directly from Figure 3 (cell-level) and Table 4 (per-class).

Comparison	$n$	$+/n$	Mean (sd) diff (pp)	95% CI (pp)	Wilcoxon $p$
Cell-level (Figure 3, balanced accuracy across 9 (feature, task) cells)
CarLab 2025 vs. da Silva Moura 2024	9	9/9	$+4.48(2.30)$	$[+3.08,+5.93]$	$0.004$
CarLab 2025 vs. USVAC 2025 Level 1	9	9/9	$+4.97(3.22)$	$[+2.99,+6.95]$	$0.004$
CarLab 2025 vs. USVAC 2025 Level 2	9	9/9	$+15.85(3.15)$	$[+14.05,+17.91]$	$0.004$
Per-underlying-class (Table 4, recall across 9 disorders)
CarLab 2025 vs. Narrow	9	8/9	$+24.89$ (19.08)	$[+13.11,+36.44]$	$0.012$
CarLab 2025 vs. da Silva Moura 2024	9	7/9	$+7.89$ (12.24)	$[+1.11,+16.11]$	$0.068$
CarLab 2025 vs. USVAC 2025 Level 1	9	3/9	$-2.22$ (14.73)	$[-10.67,+7.33]$	$0.496$
CarLab 2025 vs. USVAC 2025 Level 2	9	8/9	$+18.89$ (13.41)	$[+10.78,+27.33]$	$0.008$

“pp” $=$ percentage points; “$+/n$” is the count of paired comparisons in which CarLab 2025 had the higher value. Mean (sd) and 95% CI are computed on the paired differences (CarLab 2025 minus reference); the CI is a paired bootstrap percentile interval over $2\times {10}^5$ resamples. The Wilcoxon $p$ is the two-sided exact $p$-value of the paired signed-rank statistic under its discrete null distribution; the per-class CarLab 2025 vs. da Silva Moura 2024 comparison contains one tied pair which the Wilcoxon test drops before computing $W$. We additionally verified each row against a sign test, a paired $t$-test, and a sign-flip permutation test, all of which agreed with the Wilcoxon conclusions.