Trustworthy Evaluation of Quantum Message Passing: Isolating Circuit Contributions in Traffic Forecasting

Hybrid quantum--classical models are difficult to interpret when a quantum circuit is surrounded by trainable classical projection and readout layers, since comparable predictive performance does not reveal whether the circuit itself contributes useful information. We study this question in quantum message passing for traffic forecasting, using a variational circuit as the local aggregation operator inside a T-GCN-inspired recurrent model. On a 50-node METR-LA subnetwork, we evaluate Strongly Entangling Layers (SEL) and a road-weighted Quantum Approximate Optimization Algorithm (QAOA)-style ansatz under $R_y$, ZFeatureMap, and ZZFeatureMap encodings, at two qubit budgets and over five seeds. We isolate the circuit through matched controls that delete, freeze, or corrupt it, vary the information supplied to it, and include parameter-matched classical baselines and an exact classical surrogate for the bare $R_y$ measurement. The results show that, under the QAOA-style ansatz, removing or freezing the circuit leaves forecasting error equivalent within a pre-specified margin, while corrupting its outputs causes a large performance collapse. In contrast, training the SEL circuit materially affects accuracy, although it provides no advantage over the classical controls. ZZFeatureMap remains substantially less stable as the qubit budget increases, and finite-shot evaluation introduces error larger than the observed quantum--classical gap. These findings show that apparent hybrid-model parity can mask very different circuit contributions and motivate intervention-based evaluation as a prerequisite for trustworthy claims about hybrid quantum models.