Interpretable early warning of misinformation cascades via latent-state response dynamics

Current misinformation studies often optimize retrospective veracity classification, leaving pre-verification monitoring underdeveloped. Using the CHECKED Weibo dataset (2057 events; 3.05 million interactions), we reconstruct cascades into sliding windows and estimate response uncertainty, barrier weakness, and diffusion complexity. Because these distribution-based proxies require minimum within-window activity, latent-state estimation uses a coverage-gated subset of 1669 events, and the early-warning task uses a further nested cohort of 853 events with a fully observed 4-hour window, a subsequent 6-hour horizon, and no regime entry during the observation period. A pooled Gaussian hidden Markov model identifies a shared critical risk regime, and misinformation-linked events in the gated sample enter this regime earlier (median first-entry time 4.15 h vs. 5.37 h for verified events). Transparent regularized linear models forecast future regime entry from the first 4 h of activity, reaching an AUC/PR-AUC of 0.735/0.702 on the reference split and a mean AUC of 0.665 across repeated random splits, which we treat as the primary performance estimate; diagnostic checks rule out classic overfitting and are consistent with split-level variability. These results show that latent-state regime monitoring offers an interpretable complement to end-label classification for proactive platform triage.