The Semantic Deviation Index (SDI): A Runtime Measurement Standard for the Governance of Agentic AI in Financial Services

Abstract

Existing control frameworks in financial services share a common structural limitation: they measure what agentic systems do, not what those systems understood before they acted. The failure originates before execution begins, at the layer where meaning is resolved. It is not explicitly measured within existing governance frameworks and is not directly observable through post-execution monitoring.

This paper introduces the Semantic Deviation Index (SDI) — proposed as a candidate measurement standard for agentic AI governance in regulated financial services environments. The SDI is a runtime measurement standard that quantifies the divergence between the definition an agentic system resolves at the orchestration layer and the definition the institution has authorized. Conceptually, SDI = D(S_agent, S_authorized), where D represents a bounded, reproducible divergence measure across semantic attributes relevant to the control context.

This paper is the third in a connected body of work. Agentic Workflow Drift and Agentic Workflow Subversion (Doyle-Spare, 2026a, SSRN No. 6459612) named the failure mode and the risk surface. This paper makes both measurable. The SDI determines whether execution is governed before it occurs — a capability no existing framework explicitly measures.

The paper introduces the Deterministic Gate as the enforcement mechanism, the Semantic Audit Trail as the governance record, the Agentic Blast Radius as the containment boundary, and the Doyle-Spare Agentic Governance Model as the complete governance architecture. It establishes three governability conditions for Reasoning Baselines and demonstrates that the structural limitation of existing frameworks is categorical, not incremental: extending them to reach the reasoning layer would require observing and validating semantic resolution prior to execution, which they were not designed to perform and cannot reconstruct from post-execution artifacts.

Drawing on practitioner expertise in banking controls, model risk governance, and enterprise architecture, this paper is submitted as original practitioner-led research into an emerging and underexamined governance gap.

Keywords: Semantic Deviation Index, SDI, Doyle-Spare Agentic Governance Model, Runtime Measurement Standard, Agentic Workflow Drift, Agentic Workflow Subversion, Deterministic Gate, Semantic Control Plane, Agentic Blast Radius, Semantic Audit Trail, Candidate Measurement Standard, Reasoning Layer, Banking Controls, Model Risk Governance, Financial Services, SR 11-7, NIST AI RMF, Agentic AI Governance, Pre-Execution Validation, Semantic Alignment, Governance Latency, Authority Decay, Invisible Failure, Reconciliation Gap, Adversarial AI, Multi-Agent Systems, Control Architecture, Enterprise AI Risk, Deterministic Enforcement, Knowledge Graph, Reasoning Baselines

JEL Classification: G21, G28, G32, M15, O33, D83