Abstract
Text-free misinformation detection often turns to propagation when article or post text is unavailable. Yet public benchmarks preserve incompatible views of propagation, and graph models are rarely tested against strong summaries of the same released information. We audit summary, tabular, and graph families across nine corpora. Four event-list releases provide timestamps and event identifiers but neither retweet edges nor stable cross-cascade identities. On these releases, no evaluated graph neural network has a multiplicity-controlled accuracy advantage over a two-layer multilayer perceptron (MLP) using seven cascade summaries. Our full diagnostic configuration also has no resolved paired advantage over this MLP. Five releases provide retweet trees. We select families on validation data for the two UPFD corpora and retain exploratory test selection for the other three. On UPFD-GossipCop, validation selection picks BiGCN-style and HistGBT, with a paired accuracy difference of 0.013 (95% CI [0.005, 0.020]). Although the interval excludes zero, the difference lies inside our 0.03 equivalence margin. Input controls further localize this gain to released profile attributes rather than attachment patterns. Across generalization tests, future-test macro-F1 falls by 0.18 to 0.27 on PolitiFact and by 0.014 to 0.083 on GossipCop, while cross-corpus transfer is weak. The audit shows how released signals can govern apparent architecture gains. We therefore propose a same-information audit protocol with validation-based selection, paired inference, multiplicity and equivalence reporting, and temporal and cross-corpus evaluation.