Pretrained Graph Representations for Held-Value Counterfactual Prediction

We test whether frozen GraphGPT states support prediction at an unseen delay value. The corpus contains 690 runs and 110,449 per-second graphs. We exclude 150 ms from pretraining and readout fitting, using 67 fit pairs and 32 test pairs. Only baseline information and the proposed intervention enter the predictor. Held-150 latency MAE falls from 0.234 with intervention descriptors alone to 0.066–0.067 with raw state or bin embeddings. Raw state plus the observed baseline outcome reaches 0.064. Adding bin embeddings does not improve this estimate, and paired group-bootstrap intervals include zero. Retransmission fraction shows no demonstrated incremental benefit. Frozen states support predictive reuse within this protocol. Shared configuration groups and only 32 test pairs limit generalization and precision. Configuration matching does not identify causal counterfactual effects.