GOVERNED DECISION-INTELLIGENCE (GDI)

Abstract

Production analytics increasingly sits in the control path of consequential, often regulated decisions - which claim to pay, which applicant to approve or decline, which risk-adjustment code to submit. The model is only one component, and a technically healthy service can still be an indefensible decision-maker when data quality, feature lineage, calibrated uncertainty, explanation, human oversight, and realized outcomes are governed unevenly. This paper specifies Governed Decision-Intelligence (GDI), a framework whose unit of assurance is the decision rather than the model and whose deliverables are validation, conformance, and audit defensibility. GDI composes established instruments - the SR 11-7 model-risk lifecycle, the NIST AI Risk Management Framework, ISO/IEC management and data-quality standards, exact TreeSHAP attribution, split-conformal prediction, post-hoc calibration, and distributional drift statistics - into one testable control structure delivered as four instruments: a seven-layer reference architecture with an explicit governance spine and closed outcome loop; eight normative invariants at RFC 2119 strength with a conformance rubric; a model-risk and validation methodology; and an evidence-and-explanation model that binds every decision to the sources, versions, calibrated confidence, and deterministic explanation required to reconstruct and defend it. The decision-quality target is grounded in a construct model relating analytics use, competency, and decision quality, with testable hypotheses. The framework is portable across regulated domains and is instantiated for risk-adjustment coding, credit underwriting with adverse-action obligations, governed credit-decisioning delivery, payment-integrity analytics, and as a normative engineering standard. GDI governs how a decision is produced and certified; the run-time operation of the system that then acts on the decision is the concern of a companion production-platform architecture, with which GDI is composable.