Part 1 – How Decision Systems Learn What Matters: A Constrained Architecture for Purpose-Aligned Governance

Programme Context This preprint forms part of the research programme The Coherence Problem: How Institutions Learn, Drift, and Realign, which studies institutional decision systems as interpretive learning systems operating under conditions of complexity, scale, and delayed feedback. The programme integrates four complementary components:(1) architecture — the formal structure of decision-system learning,(2) mechanism — translation drift as a structural source of misalignment,(3) measurement — methods for observing translation coherence, and(4) design — governance as interpretive maintenance in AI-mediated environments. Together, the papers examine how organisations determine what matters, how meaning becomes encoded in governance artefacts, how translation drift arises as intent moves across governance layers, and how institutions can observe, maintain, and deliberately realign interpretive coherence over time. Preprint Description This paper presents the architectural foundation of the programme. It introduces AI-Augmented Impact Frames, a formally specified, constrained, closed-loop architecture for studying how organisational decision systems learn over time. The framework treats the decision system—rather than individual decisions or decision-makers—as the primary unit of analysis. At the centre of the architecture is the Operating Spine, a structured chain linking purpose, capabilities, value drivers, strategy, portfolios, and signals in a closed learning loop. The framework combines: • a formal decision-learning architecture (the Operating Spine)• constrained AI-supported interpretation that preserves human decision authority• psychometrically grounded longitudinal measurement using Item Response Theory (IRT) Rather than optimising or automating decisions, the architecture is designed to make decision logic explicit, measurable, and falsifiable. AI functions as interpretive support and is structurally prohibited from recommendations or decision authority. This paper defines the boundary conditions, failure modes, and falsifiability criteria required for empirical evaluation of decision-system learning under conditions of delayed feedback and high uncertainty, and provides the structural foundation for the subsequent papers on mechanism, measurement, and design. Version History Version 1.01 clarified terminology, strengthened falsifiability statements, and improved conceptual definitions without changing the underlying architecture. No empirical datasets are associated with this version. Measurement instruments, code scaffolds, and empirical studies will be released in subsequent linked records. Version 1.03 update: This version restored the intended architectural manuscript file. Version 1.02 contained a mistakenly uploaded file that did not reflect substantive changes to this record. No conceptual, methodological, or structural content was altered relative to Version 1.01. Version 2.00: This release consolidates the manuscript within the full research programme structure. Cross-paper terminology has been harmonised, the unit-of-analysis statement has been standardised across the series, and reference architecture has been aligned. No changes have been made to the formal decision-learning architecture, measurement logic, boundary conditions, or theoretical claims. Version 2.01: consolidates the manuscript within the full research programme structure. Cross-paper terminology has been harmonised, titles and references have been aligned with the programme statement, and internal cross-references have been updated. No changes have been made to the formal decision-learning architecture, measurement logic, boundary conditions, or theoretical claims. Empirical studies, measurement instruments, and field applications are in preparation and will be released in subsequent linked records.