Abstract
Temporal link prediction is usually scored by a single outer-test average, which treats predictive risk as exchangeable across dyads. Contact and online-learning streams violate that assumption: recurrent motifs coexist with sparse, weakly recurrent edges. We study whether the relative out-of-sample risk of snapshot message passing is heterogeneous across local weighted-degree-entropy states, and whether an entropy-conditioned router yields selective rather than uniform risk reductions. On SocioPatterns school/workplace contacts, BenchTemp massive open online course (MOOC) and contact-like panels, and synthetic temporal graphs, we compare a calendar–recency logistic baseline, protocol-matched temporal graph neural network (GNN) controls, an ungated temporal graph convolutional network (GCN), and an entropy-gated dual-head model under a chronological 50/25/25 split. Normalized local weighted-degree entropy serves as both an evaluation stratum and a soft gate between message-passing and calendar–recency logits. In a fixed medium protocol (12 units), gating improves mean outer-test log-loss relative to the ungated GCN on school ( ) and workplace ( ), with school gains concentrated in the highest-entropy quintile (mean ). These gains are relative to a weaker ungated GCN; a calendar–recency logistic baseline often remains competitive or better in absolute log-loss. Heterogeneous-exponential synthetics favor gating, whereas memory-free and homogeneous-exponential controls can produce harm; multi-seed and transfer screens bound stability. The contribution is diagnostic: entropy-conditioned errors matter when risk is uneven, but the effect is neither universal across domains nor stable under all generators.