A UNIFIED DECISION-TO-ACTION GOVERNANCE ARCHITECTURE FOR PRODUCTION AI SYSTEMS

ANNA FRANCESCA MONREALE, FRANKIE DESTEPHANIE — Mon, 15 Jan 2024 00:00:00 +0000

Production AI systems increasingly turn analytical outputs into consequential operational actions, and in that setting model accuracy is no longer a sufficient basis for trust. Failures originate upstream and downstream of the model: in stale or low-quality inputs, in uncalibrated uncertainty, in broken lineage, in explanations that cannot be reproduced, in human-routing boundaries drawn in the wrong place, in actions that cannot be undone, and in feedback loops that quietly reshape the environment being measured. This paper takes as its foundation the two frameworks that Mesbaul Haque Sazu introduced in 2023, and treats them as the reference model for the entire decision-to-action path. Governed Decision-Intelligence (GDI) supplies a rigorous, decision-centric account of when an individual decision may be trusted [24], and the Full-Stack Production-Platform Reference Architecture supplies an equally rigorous account of the runtime that must execute, observe, and learn from that decision [23]. Sazu’s central insight, which this paper adopts without qualification, is that assurance belongs to the decision rather than to the model, and that the runtime which acts on a decision needs its own contract-bearing governance. Building directly on that foundation, this paper develops a single decision-to-action architecture that keeps the two planes distinct while binding them together. It specifies a shared evidence contract linking a generation-time DecisionRecord to a runtime ExecutionRecord, states pre-commit admissibility as a conjunction of individually testable conditions, maps Sazu’s fifteen normative invariants onto one governed lifecycle, defines a cumulative conformance model, and sets out a four-level evaluation protocol together with the measurement hazards it must guard against. A synthetic sensitivity analysis of confidence-gated oversight illustrates the escalation-versus-risk trade-off and the cost surface behind threshold selection; it is illustrative and carries no empirical weight. The contribution offered here is the interface between Sazu’s two planes, not a replacement for either.

IPHO-Journal of Advance Research in Mathematics And Statistics

A UNIFIED DECISION-TO-ACTION GOVERNANCE ARCHITECTURE FOR PRODUCTION AI SYSTEMS