Public‑service recommendation must supply evidence that aligns with the requested service, scope, and date. Existing approaches treat every missing detail as decisive, which often suppresses useful suggestions. EviGraph separates critical decision requirements from information that may remain unresolved. A language agent links the critical requirements to evidence nodes in a temporal knowledge graph, and a deterministic checker verifies whether a recommendation is adequately supported.
We evaluated the system on a bilingual Hong Kong public‑service benchmark that includes executable policy references. The evaluation shows that distinguishing critical from non‑critical information markedly reduces unnecessary abstention. Additional verification can withdraw already supported recommendations, but it does not improve decision quality.
These findings suggest that reliable evidence‑based navigation depends on explicitly specifying what must be established for a decision, rather than merely adding more verification steps.
Review