Experiment Note
Bootstrap 95 percent confidence intervals were computed for the main held-out metrics. Shielded SafeGraph achieved mean RMS 1.3381 across the broader aggregate with a reported 95 percent interval [0.9650, 1.7300], and trust AUC approximately [0.9913, 0.9941]. Unshielded SafeGraph showed a nearly identical trust AUC interval, confirming that the shield changed control intervention rather than trust-classification quality.
Paired Wilcoxon comparisons were run over matched seed-attack pairs, with Holm correction at alpha = 0.05. On attacked scenarios, comparisons against raw CACC, chi-square CACC, residual trust, and unshielded SafeGraph were significant for the reported principal metrics. There were 180 matched pairs in the main attacked comparisons.
Technical Essay
The statistical procedure was selected because each method was evaluated on the same stochastic seed and attack condition. Pairing removes part of the between-run variance and directly tests whether the proposed method improves the outcome under equivalent disturbances.
The confidence intervals and paired tests were added to reduce dependence on a single averaged score. They did not turn the study into a formal safety proof, but they made the empirical comparison much harder to dismiss as an artifact of one random seed.