Evaluating Institutional Decision Reconstruction in AI Governance

Abstract Public inquiries, courts, supreme audit institutions, and regulatory investigations have documented cases in which authorised reviewers were unable to reconstruct decision- essential institutional information. They establish that such failures occur in consequential governance settings and motivate a specific methodological question: how can institu- tional decision reconstruction be evaluated in a scientifically reproducible manner across heterogeneous governance environments? Existing research addresses complementary questions concerning decision representation, records, provenance, traceability, assurance, organizational processes, institutional facts, and document review. The residual problem addressed here is not the absence of those constituent capabilities, but the specification of a common procedure for evaluating the institutional conclusions that independent reviewers derive from heterogeneous governance evidence. This paper develops the foundations of a reproducibility-oriented evaluation method for institutional decision reconstruction. The contribution is methodological rather than theoretical. The paper neither introduces a new theory of organizational decision-making nor proposes a new governance framework or enterprise ontology. Instead, it specifies: (i) a formal reconstruction-task definition; (ii) a relational evaluation unit that indexes performance to reviewer, record arrangement, review purpose, and resource budget; (iii) externally adjudicated, corpus-relative reference conclusions; (iv) a multi-label reconstruc- tion error taxonomy; (v) a controlled factorial design separating representation effects from elicitation-instrument effects; and (vi) a staged validation programme that licenses progressively stronger classes of scientific inference. The method distinguishes ordinary inaccuracy from critical institutional errors, false closure, missed closure, bounded closure, calibration error, and revision behaviour. A controlled synthetic scenario, Northbridge, illustrates how the task, reference conclusions, error taxonomy, factorial design, and evidence-progression procedure can be operationalized. It provides no empirical validation. The manuscript adopts a conservative epistemic position. It does not claim that institutional decision reconstructability is a distinct latent construct, that the proposed method is already empirically validated, or that any particular record arrangement, reviewer type, governance framework, or software implementation performs better. Reliability, discriminative power, external adjudication, cross-case robustness, field correspondence, and practical utility remain explicit empirical requirements. By prioritizing measurement before theory, controlled evaluation before generalization, and reproducibility before explanation, the proposed method provides a transparent foundation for cumulative research on institutional decision reconstruction, including prospective applications in AI governance.