A preprint reports strong low-data results for speech-act-informed models forecasting online conversation derailment, but mixed rankings across domains.