EviGraph Uses Temporal Evidence to Support Public-Service Recommendations

Summary

Public-service recommendation systems need evidence that matches a requested service, geographic scope, and date, but treating every unresolved detail as a reason to abstain can suppress useful guidance. EviGraph addresses this trade-off by separating critical decision requirements from information that may remain unresolved. A language agent maps those requirements to supporting evidence in a temporal knowledge graph, and a deterministic checker tests whether the resulting recommendation is actually supported. The system was evaluated on a bilingual Hong Kong public-service benchmark containing executable policy references. The evaluation found that distinguishing essential requirements from nonessential missing details reduced unnecessary abstention. It also found a limitation: adding more verification could cause the system to withdraw recommendations that were already supported without improving decision quality. The authors therefore argue that reliable evidence-based navigation depends on specifying what must be established for a particular decision, rather than simply increasing the amount of verification.