Decision Demonstrability: Why Auditable Outputs Are Necessary but Not Sufficient

This paper introduces two concepts for the governance of automated decision systems: Decision Demonstrability: the ability to publicly reconstruct, in an individual case, why a given output qualifies as an instance of the governing decision category under which it is issued. Decision Identity Failure: a condition in which a system output is formally categorized as a specific type of decision, but the derivation that would make the output an instance of that category is absent or not independently reconstructable. The paper argues that existing governance frameworks, including NIST AI RMF 1.0 and ISO/IEC 42001:2023, converge on auditability, explainability, and regulatory compliance as primary evaluative criteria, but that these criteria are necessary and not sufficient. A decision system may satisfy all three while failing to produce outputs that are demonstrable instances of the governing category under which they are issued. Decision Identity Failure is shown to be orthogonal to outcome correctness, regulatory compliance, and explainability. Three necessary conditions are defined. The concept is applied to automated medical necessity determinations (UnitedHealth Group / nH Predict) and generalized across five domains: credit decisions, recidivism risk assessment, autonomous weapons systems, administrative decisions, and agentic AI systems. Four implications for governance framework design are derived, including the treatment of the governing category as a governance parameter subject to re-authorization when normative requirements change. Framework references: DIP-CORE-1.0 / GCCL v0.1 (DOI: 10.5281/zenodo.18362037)Related: DIP Audit #6 - Decision Identity Failure in Automated Medical Necessity Determinations (DOI: 10.5281/zenodo.19744184) decision governance · automated decision systems · decision demonstrability · categorical validity · AI audit · Decision Identity Failure · AI governance · medical necessity · accountability