EVIDENCE-TO-DELIVERY PROTOCOLS FOR CONSISTENT MULTI-ARTIFACT DATA ENGINEERING

LLM pipelines for data engineering can generate individually plausible artifacts—models, schemas, mappings, and execution plans—without ensuring they form a coherent delivery package. This gap motivates an evidence-to-delivery framework that separates and tests two mechanisms: candidate generation and candidate selection. We propose PlanGate, which shares one relation plan across eight requested artifacts and delegates DDL admission to a reference-free deterministic predicate. To isolate how relational evidence affects each protocol, we introduce a factorial intervention over foreign-key declarations and meaningful column names, combined with matched planned/unplanned streams for generation effects and fixed-pool replays for selection effects. We also present DE-VeriBench, 306 cases across seven source packages, evaluating structural fidelity, parsing, execution, and package consistency as distinct outcomes. Across seven models from four model families, PlanGate achieves higher mean foreign-key topology recovery than Single-shot for every model. Crucially, the joint analysis distinguishes a better structural proposal from executable DDL and from a fully verified package. This framework gives data engineers an explicit basis for deciding what information to preserve, where to allocate inference, and which properties must be checked before submission.