Learning Which Route to Trust: Stability-Aware Bayesian Next-Hop Selection for Infrastructure-Independent Disaster Messaging

When cellular sites lose power and backhaul after a disaster, short text mes- sages must hop between phones, shelters, and regional gateways over inter- mittent, low-rate links. Choosing the next hop is then a prediction problem: which neighbor is most likely to deliver the message before it expires? We present STAR (Stability-aware TAR routing), a fully local next-hop policy for the Taiwan Resilient Communication Network that (a) predicts route stability Ψ, the probability of on-time delivery, from discounted Bayesian posteriors over link frame success and availability combined with loop-free, authenti- cated distance-vector advertisements; (b) decides with a priority-aware rule that is risk-averse and replicates distress messages, yet explores with Thomp- son sampling on non-urgent traffic; and (c) learns from outcomes through fragment acknowledgments, signed end-to-end receipts treated as falsifiable deadlines, and overheard forwarding, which together calibrate trust in each neighbor’s route claims. In a discrete-event simulator with LoRa time-on-air and airtime-accounted control traffic, evaluated on 20 held-out seeds of a syn- thetic two-region disaster network, STAR delivered 97.7% of messages and all distress messages, against 89.6% for a loop-free ETX distance-vector baseline and 90.5% for PRoPHET, with 44% lower median latency than the former. After a simulated aftershock it delivered 92.8% versus 82.5% and 84.9%. Un- der route poisoning it retained several times the delivery of an unprotected distance-vector protocol, although replication-based protocols remained more robust. Gains vanish at saturation. All results are synthetic.