Progress Update
The evaluation was structured around a controller ladder rather than a single proposed-versus-raw comparison. The principal baselines became:
- Raw CACC using communicated predecessor states directly.
- No-V2V or sensor-only ACC.
- Chi-square reliability-weighted CACC.
- Residual-trust CACC.
- EdgeMLP trust CACC.
- Unshielded SafeGraph-CACC.
- Shielded raw CACC.
- Shielded EdgeMLP.
- Shielded SafeGraph-CACC.
- Oracle-trust CACC as a non-deployable upper bound.
Technical Essay
This baseline design allowed the experiment to answer multiple questions:
- Does V2V create a meaningful attack surface? Raw CACC versus sensor-only ACC.
- Does learned trust improve over conventional statistical consistency? SafeGraph versus chi-square.
- Does graph context matter beyond per-edge features? SafeGraph versus EdgeMLP.
- Does physics residual weighting matter? Learned graph trust versus simpler residual trust.
- Is learned trust alone sufficient? Unshielded versus shielded SafeGraph.
- Is the shield doing all the work? Shielded raw CACC versus shielded SafeGraph.
- How far is the learned method from perfect attack knowledge? SafeGraph versus oracle trust.
The addition of shielded raw CACC was especially important. Without it, reviewers could reasonably argue that any safety improvement came only from the runtime shield and that the graph-trust component was unnecessary. The ablation made the architecture auditable rather than monolithic.