Evaluates decision-making accuracy in agentic enterprise systems, suggesting improved evidence support.
Enterprise agents can cite documents while still making unsupported decisions: a source can be stale, outside the caller's permitted scope, privacy-incompatible, unrelated to the asserted claim, or disconnected from the action subsequently taken. This paper proposes CGE, a Citation-Guided Evaluation framework for agentic enterprise systems. CGE represents claims, retrieved evidence, source authority, tool calls, privacy transformations, policy decisions, and operational outcomes as a typed citation graph. It evaluates not only citation presence, but support entailment, authority coverage, provenance completeness, authorized use, influence propagation, and outcome calibration. CGE is applied to simulated documentation maintenance, policy-explanation retrieval, synthetic-data governance, and pipeline-observability tasks. Relative to provenance logging without graph verification, CGE increases supported-action precision from 0.79 to 0.93, improves complete evidence-chain coverage from 0.68 to 0.91, reduces stale-influence acceptance from 8.9% to 2.1%, and rejects all simulated unauthorized evidence-use events.
No takes yet. Share an insight, caveat, or question.
Malempati et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: