Agentic AI Semantic Deviation Index - A Reference Specification with a Working Python Implementation

The Agentic AI Semantic Deviation Index is a runtime measurement construct for Agentic AI governance in regulated financial institutions. It quantifies Runtime Semantic Divergence, the single-decision condition in which an agent’s runtime operational interpretation diverges from the institution’s authorized Reasoning Baseline. The construct is defined in Doyle-Spare (2026b) and given its operative form in the Semantic Control Plane architecture published as Doyle-Spare (2026), where the Deterministic Gate consumes the Index to determine whether Semantic Authorization may be emitted before an agent acts. That published architecture incorporates by reference a normative mathematical specification covering the formal definitions of the component scores, the calibration discipline, and the threshold governance procedures. This document provides the technical specification for computing, representing, calibrating, classifying, and testing the Semantic Deviation Index. It defines the component contracts, canonical aggregation, numerical representation, calibration discipline, threshold-zone classification, required measurement evidence, interoperability conditions, and conformance requirements. The Semantic Control Plane, Deterministic Gate, authorization lifecycle, failure routing, and Semantic Audit Trail remain governed by Doyle-Spare (2026) and are referenced only where they constrain the SDI interface. The formal properties every component method must satisfy are normative here. The methods themselves are the subject of future illustrative work and validation. A separately versioned reference implementation and conformance suite may demonstrate implementation of the finalized specification and provide the executable vectors against which conformance is asserted. Neither artifact defines or extends the normative architecture.