Experiment Note
The sensitivity suite was extended across platoon size N = 4, 6, 8, 12, and 16. Shielded SafeGraph maintained zero crash penetration across the tested sizes. Mean RMS decreased from 2.6674 at N = 4 to 0.8581 at N = 16 under the reported aggregation; minimum gaps remained roughly 12.1-13.1 m. The N = 6 condition produced a small mean unsafe count of 0.08 while still showing zero collisions and zero crash metric. This exception was retained rather than hidden.
Headway sensitivity was tested at h = 0.4, 0.6, 0.8, 1.0, and 1.2 s. Shielded SafeGraph remained collision-free and crash-free across all tested settings. As expected, minimum gap increased with headway, while RMS tracking error rose modestly at conservative headways.
Delay stress used 0, 1, 2, 3, and 5 simulation steps, corresponding approximately to 0-0.5 s in the configured lightweight experiment. Trust AUC remained near 0.99 through moderate delay but fell to 0.8420 at five steps. RMS error increased sharply to 13.8890 at the most severe delay. The shield still prevented collisions and crash penetration, but small unsafe-gap counts appeared at three and five steps.
Runtime benchmarking showed Shielded SafeGraph at approximately 1.1410 ms mean, 1.1802 ms p95, 1.1950 ms p99, and 1.1987 ms maximum per simulator step in the Python benchmark. It remained below the nominal budgets for 10 Hz, 20 Hz, and 50 Hz loops.
An out-of-distribution clean sanity check showed all methods collision-free. SafeGraph preserved clean CACC behavior with RMS 0.2531 versus 0.2504 for raw CACC and trust AUC 0.9731.
Technical Essay
These diagnostics exposed the actual boundary of the contribution. The method was robust to larger platoons and different headways, but severe temporal staleness remained difficult. The shield protected against the worst physical outcomes while tracking degraded. This distinction prevented the manuscript from presenting "zero collisions" as proof that the entire control problem had been solved.