Recurring analytics reviews often require narratives that follow stable organizational conventions, including what to explain, how to decompose changes, and which drivers are appropriate for attribution. Large language models can generate fluent summaries from tables and dashboards, but prompt-based workflows are difficult to audit because the source of errors in scope, decomposition, attribution, or evidence is often unclear. We present Ontology-Constrained Narrative Interaction (OCNI), a framework that represents analytic constraints as explicit interaction controls. Instead of relying on prompt rewriting, OCNI guides generation through ontology-backed policies that encode valid decompositions and attribution paths. Given an analyst-selected scope and policy, the system retrieves a relevant ontology subgraph, constructs a constrained narrative plan, and generates a report grounded in explicit entity-metric-time evidence. To support verification and repair, OCNI emits a trace artifact called the Claim Ledger, which links each narrative claim to its supporting evidence and ontology reasoning path. When evidence is insufficient or an attribution is inadmissible, the system surfaces explicit constraint violations rather than speculative explanations, enabling targeted repair through scope or policy adjustments. We instantiate OCNI in a recurring analytics reporting pipeline and evaluate it through expert review and automated traceability checks, including an ablation against prompt-only generation. Results show that ontology constraints improve narrative traceability and structural consistency while enabling transparent failure recovery.
Chadalavada et al. (Fri,) studied this question.