Compare staleness–cost trade-offs across grounding architectures

Determine whether staged, hierarchical, or live-only grounding architectures trade staleness against cost differently when evaluated with the ChurnBench benchmark.

Background

ChurnBench measures freshness errors in agentic systems operating over changing enterprise data, distinguishing answers that were once correct but became stale from answers that were never correct. The reported experiments use a single grounding architecture with tiered refresh scheduling and do not compare alternative architectures.

The paper identifies comparative evaluation of staged, hierarchical, and live-only designs as a direct unresolved question. Such a comparison would assess whether these architectural choices produce different trade-offs between data freshness and operational cost under controlled drift, rather than merely comparing refresh policies within one implementation.