The Fabric Data Platform Kit: Spec-Driven, Parity-Gated Analytics Delivery in the Agentic Development Lifecycle

Abstract

This third paper in the Agentic Development Lifecycle series takes its argument from architecture to artifact. It describes the Delivery Kit—the installable, versioned unit in which Evoke packages a complete Program-layer capability—and examines one flagship kit in depth: the Source-to-Fabric Data Platform Kit, which arms a single AI-augmented engineer to take raw source data to a governed Microsoft Fabric lakehouse, medallion ETL, star schema, Direct Lake semantic model, and Power BI report. The kit is organized as a six-phase, spec-driven playbook in which every phase produces a reviewable artifact, three named human gates preserve accountability, and nothing is called "done" until a four-point parity contract—row parity, measure parity, grain integrity, and semantic parity—passes under signed tolerances. We show how five converged lanes are carried by one engineer and why the validation gate, rather than the generated code, is the kit's center of gravity. Evidence from a factory-bookings engagement demonstrates both paths: a clean run reconciling to the cent, and a planted silver-layer defect that a plausible-looking report would have hidden but the parity gate caught. The paper closes with the honest limits of the approach—assisted rather than push-button delivery, dependence on the quality of the specs, and the risks of tolerance drift.