From Inference to Consequence: Operational Integrity as a System-Level Benchmark for Operational AI

Artificial intelligence is increasingly moving from information generation toward operational participation, where AI systems invoke tools, modify enterprise data, execute workflows, and create financial, legal, or organizational consequences. In such environments, model capability and task success alone provide only a partial account of system quality. This paper introduces Operational Integrity as a system-level evaluation property and presents the conceptual and methodological foundations of the Operational Integrity Benchmark (OIB). The framework distinguishes probabilistic inference from an Operational Decision and evaluates whether externally effective consequences arise only through authorized, verified, governed, auditable, and reproducible operational processes. The proposed methodology reconstructs governance-relevant operational semantics from heterogeneous Runtime Evidence. It derives an Operational Event Trace, reconstructs an implementation-independent Canonical Operational State, projects a scenario-specific Observed Operational State, and validates the resulting behavior against independently specified Ground Truth. The OIB treats Commit, Block, Escalate, Defer, and Non-Decision as potentially correct operational outcomes and derives standardized Operational Integrity metrics only after completed validation. The framework is technology-neutral and does not prescribe a specific runtime-governance architecture.