Continual Adaptation Under Missing Propagation Structure: Mechanisms and Limits of a Gated Semantic-Topological Network for Rumour and Harmful-Content Detection

We study sequential adaptation from rumor detection on PHEME, where reply structure is observed, to harmful-content flagging on USE24-XD, a 2024–2025 corpus released without propagation structure. A gated network fuses a BERTweet representation of the source post with a graph representation of the reply cascade, and Elastic Weight Consolidation (EWC) limits forgetting during fine-tuning. With a fixed split, 20 initialization seeds, matched training protocols and validation-only selection of λ, the gated model is no better on PHEME than a model of the source post alone, and neither reply structure nor learned gating improves on fixed fusion; under leave-one-event-out evaluation, however, the graph model generalizes best. Naive fine-tuning reduces PHEME balanced accuracy to below chance (0.481), whereas EWC retains 0.561 (+0.080, 95% CI [0.063,0.098], all 20 seeds): about a fifth of the pre-adaptation margin over chance, at a cost of 0.017 on the new task. Uniform anchoring of equal strength retains as much but prevents adaptation. EWC, not the new data, shifts the gate towards the graph branch, which on edgeless graphs is exactly a multilayer perceptron: the penalty freezes the semantic projection while the graph branch adapts.