Submission Record
Submitter: Maureen Doyle-Spare
Document commented on: NIST AI Risk Management Framework Playbook — MEASURE function, subcategories MEASURE 3.2, MEASURE 3.1, MEASURE 1.1 and MEASURE 2.5
Date submitted: April 16, 2026 (2026-04-16)
Key IP in Submission — Agentic Governance Model
Semantic Deviation Index (SDI) — the subject of the submission, proposed as a reference measurement standard for the MEASURE function for agentic AI systems operating across multi-system enterprise environments. The runtime measurement of divergence between the agent’s operational interpretation and the institution’s authorized Reasoning Baseline at the orchestration moment. Composite normalized to a 0 to 100 scale and classified against four threshold zones that determine the Gate decision. Formalized in §8. 10.2139/ssrn.6531238 SDI code created and formalized in maureendoylespare/agentic-ai-semantic-deviation-index — Semantic Deviation Index reference specification and exact-rational Python implementation for pre-execution governance of the reasoning layer of agentic AI.
Agentic Workflow Drift (AWD) — named in the prior Playbook filing this submission follows on from. The cumulative institutional operating-state degradation that emerges when unauthorized resolutions are repeatedly executed at the reasoning layer without runtime governance. AWD operates at the institutional-state timescale, distinguished from Runtime Semantic Divergence (RSD), the execution-stage condition at a single agentic decision. Formalized in §3.1 and Doyle-Spare (2026a). 10.2139/ssrn.6459612
Agentic Workflow Subversion (AWS) — named in the prior Playbook filing this submission follows on from. The institutional condition in which an agentic workflow proceeds under an unauthorized resolution. Agentic Workflow Subversion may arise through accumulated, unmitigated AWD or through intentional manipulation of runtime semantic resolution. Formalized in §3.1 and Doyle-Spare (2026a). 10.2139/ssrn.6459612
Reasoning Layer — the reasoning-layer measurement gap the prior filing identified. The governance surface where an agent resolves operational meaning before execution. It sits between model output and execution and is evaluated through governable runtime representations rather than latent cognition or internal chain-of-thought. 10.5281/zenodo.20749051
Incorporation requested
MEASURE 3.2: incorporation of the Semantic Deviation Index in the next Playbook update cycle
MEASURE 3.1: incorporation of the Semantic Deviation Index in the next Playbook update cycle
MEASURE 1.1: incorporation of the Semantic Deviation Index in the next Playbook update cycle
MEASURE 2.5: incorporation of the Semantic Deviation Index in the next Playbook update cycle
Prior NIST submissions disclosed in this filing
Six prior NIST submissions on this framework were filed before the March 31, 2026 deadline, across:
AI RMF feedback channel; NCCoE; COSAiS and NISTIR 8605D; NIST AI 800-2; AI RMF Engage and Crosswalks; and the CAISI Finance Sector Listening Session.
Named acknowledgment was received from NIST and CAISI in March 2026.
Related submission
This filing follows a prior Playbook submission of March 2026 through the same channel, which identified the reasoning-layer measurement gap and introduced the Semantic Deviation Index as a candidate construct. The present submission delivers the formal specification.
Index terms
regulatory policy; AI policy; regulatory engagement; AI Risk Management Framework; AI RMF Playbook; MEASURE function; MEASURE 1.1; MEASURE 2.5; MEASURE 3.1; MEASURE 3.2; reference measurement standard; runtime measurement; multi-system enterprise environments; CAISI; NCCoE; COSAiS; NISTIR 8605D; NIST AI 800-2; autonomous systems; multi-agent systems
Public Comment Submission Details
Transmitted April 16, 2026 to the NIST AI RMF Playbook rolling feedback channel. Reproduced as sent
Dear AI RMF Playbook Team,
I am submitting the attached document as a standalone reference contribution to the AI RMF Playbook rolling feedback channel. The frameworks referenced in this submission are the independent research of the submitter, developed in a personal capacity and not associated with any employer. The submission proposes the Semantic Deviation Index (SDI) as a reference measurement standard for the MEASURE function, specifically for agentic AI systems operating across multi-system enterprise environments. It requests incorporation under MEASURE 3.2, MEASURE 3.1, MEASURE 1.1, and MEASURE 2.5 in the next Playbook update cycle.
This submission is a follow-on to a prior Playbook filing submitted in March 2026 via this channel (Agentic Workflow Drift and Agentic Workflow Subversion: A Structural Gap in the NIST AI RMF Playbook, SSRN Working Paper No. 6459612), which identified the reasoning-layer measurement gap and introduced the SDI as a candidate construct. The attached paper delivers the formal specification. Both papers are available on SSRN:
- SSRN No. 6459612 (AWD/AWS Risk Taxonomy): https://ssrn.com/abstract=6459612
- SSRN No. 6531238 (SDI Measurement Standard): https://ssrn.com/abstract=6531238
Six prior NIST submissions on this framework were filed before the March 31, 2026 deadline across AIframework@nist.gov, NCCoE, COSAiS/NISTIR 8605D, NIST AI 800-2, AI RMF Engage/Crosswalks, and the CAISI Finance Sector Listening Session. Named acknowledgment was received from NIST/CAISI (March 2026).
Thank you for the ongoing work on the Playbook.
Best regards,
Maureen Doyle-Spare
Researcher, AI Governance Controls
linkedin.com/in/maureendoylespare
The Semantic Deviation Index (SDI)
A Measurement Standard for the MEASURE Function in Agentic AI Systems
AI RMF Playbook · Reference Contribution
Submitted by: Maureen Doyle-Spare
Role: Independent Practitioner and Researcher, AI Governance and Banking Controls
Status: Independent practitioner research, developed in a personal capacity and not associated with any employer.
LinkedIn: linkedin.com/in/maureendoylespare
Date: April 2026
Channel: AIframework@nist.gov - Rolling Feedback
Paper 1: Doyle-Spare (2026a). Agentic Workflow Drift and Agentic Workflow Subversion: A New Risk Taxonomy for the Governance of Agentic AI in Financial Services. SSRN Working Paper No. 6459612. https://ssrn.com/abstract=6459612
Paper 2: Doyle-Spare (2026b). The Semantic Deviation Index (SDI): A Runtime Measurement Standard for the Governance of Agentic AI in Financial Services. SSRN Working Paper No.
6531238. https://ssrn.com/abstract=6531238
Prior Submission:
Doyle-Spare (March 2026). Agentic Workflow Drift and Agentic Workflow Subversion: A Structural Gap in the NIST AI RMF Playbook and a Pre-Standardization Input for Financial Services. Submitted via AIframework@nist.gov.
1. Purpose of This Submission
This submission is a standalone reference contribution to the AI RMF Playbook rolling feedback channel. It proposes the Semantic Deviation Index (SDI) as a reference measurement standard for the MEASURE function, specifically for agentic AI systems operating across multi-system enterprise environments.
A prior submission filed in March 2026 via this channel identified a structural gap in the Playbook’s MEASURE function: the absence of any measurement standard for the reasoning layer of agentic AI systems - the orchestration layer where agents resolve meaning across enterprise platforms before execution proceeds. That submission (Recommendation 2: MEASURE 3 - Runtime Observability for Agentic Systems) drew a critical distinction that frames this submission: existing Playbook guidance supports reasoning observability - governing the artifacts of reasoning, what the system did, what it produced, and how it behaved. It does not yet require governance of what the system understood prior to execution, or whether that understanding remained within authorized bounds.
The failure mode that Agentic Workflow Drift describes originates prior to execution at the reasoning layer and is not observable through post-execution monitoring. The SDI is the measurement construct designed to reach it. The SDI paper (2026b) delivers the formal specification the March submission called for. This submission provides the specific MEASURE subcategory mapping that converts the March proposal from a broad category recommendation into a targeted, subcategory-level reference contribution.
2. The Measurement Gap This Submission Addresses
The AI RMF MEASURE function has no measurement standard for the reasoning layer of agentic AI systems. All existing MEASURE subcategories address outputs, behaviors, statistical distributions, or design-time properties. None addresses the semantic state an agentic system constructs at the orchestration layer - the meaning it resolves from inputs across contributing enterprise systems - and whether that meaning is aligned with what the institution explicitly authorized before execution proceeds. This gap is not a lack of observability, but a lack of measurement at the point of semantic resolution prior to execution.
Semantic deviation in agentic systems is not a performance gap. It is a leading indicator of governance risk.
As SDI increases, the probability of execution against unauthorized definitions increases - even when no statistical deviation, control failure, or output anomaly is observable.
The system performs correctly against a definition that has already shifted. The deviation is only visible at the reasoning layer, before execution.
No existing MEASURE subcategory is designed to reach it.
Without a measurement construct at that layer, the MANAGE function cannot operationalize intervention - no signal exists to trigger supersession, escalation, or containment.
3. Proposed Reference Contribution by MEASURE Subcategory
Each subcategory below is quoted verbatim from NIST AI 100-1. Each mapping follows a consistent structure: what is being measured, why current metrics fail to reach it, what the SDI measures and how it is computed, and what governance artifact it produces.
MEASURE 3.2 - “Risk tracking approaches are considered for settings where AI risks are difficult to assess using currently available measurement techniques or where metrics are not yet available.”
What is being measured:
The divergence between the semantic state an agentic system resolves at the orchestration layer and the semantic state the institution has explicitly authorized - measured before
execution proceeds, at the orchestration decision node where meaning is constructed, not where outputs are produced.
Why current metrics fail:
Every existing AI risk metric operates on outputs, behaviors, or statistical distributions. Semantic deviation produces none of these signals. The system performs correctly against a definition that has already shifted. The deviation is structurally invisible to all post-execution monitoring approaches currently referenced in the Playbook.
What the SDI measures and how it is computed:
The SDI requires three observable components at each orchestration decision node: (1) the resolved semantic state (Sₐ), captured as a structured, machine-representable semantic state representation of the agent’s interpreted control terms at execution time - derived directly from the agent’s reasoning trace, decision context, and retrieved system definitions at the orchestration node; (2) the authorized semantic baseline (Sₒ), defined through policy-linked ontological mappings or controlled vocabularies established by the institution’s governance function; and (3) a divergence function D, computed using institution-defined, governed methods - such as rule-based equivalence or embedding-based similarity - applied consistently within a defined control context.
The SDI requires three observable components at each orchestration decision node: (1) the resolved semantic state (Sₐ), captured as a structured, machine-representable semantic state representation of the agent’s interpreted control terms at execution time - derived directly from the agent’s reasoning trace, decision context, and retrieved system definitions at the orchestration node; (2) the authorized semantic baseline (Sₒ), defined through policy-linked ontological mappings or controlled vocabularies established by the institution’s governance function; and (3) a divergence function D, computed using institution-defined, governed methods - such as rule-based equivalence or embedding-based similarity - applied consistently within a defined control context.
SDI = D(Sₐ, Sₒ)
The SDI is measured at each decision node and recorded as part of the Semantic Audit Trail, enabling longitudinal tracking of semantic alignment across workflow executions.
What governance artifact it produces:
The Semantic Audit Trail - a governance record of the semantic state the agent resolved at each orchestration checkpoint, the authorized baseline against which it was compared, and the SDI score at that node. This is not an execution log. It is a pre-execution governance record: what the system understood before it acted, and whether that understanding was within authorized tolerance. A Deterministic Gate can operationalize the threshold: when the SDI score exceeds governed tolerance, execution is halted and escalated for human review before proceeding.
MEASURE 3.2 explicitly addresses settings where metrics are not yet available. The SDI is a proposed metric for this setting. It is bounded, reproducible, auditable, and capable of consistent implementation across systems operating under a shared governance framework. Implementation complexity will vary based on system architecture and governance maturity. The measurement construct itself remains consistent across deployments.
MEASURE 3.1 - “Approaches, personnel, and documentation are in place to regularly identify and track existing, unanticipated, and emergent AI risks based on factors such as intended and actual performance in deployed contexts.”
What is being measured:
Semantic alignment drift across successive executions - the longitudinal pattern of SDI scores at each orchestration decision node, tracked against the authorized Reasoning Baseline over time.
Why current metrics fail:
Semantic deviation in multi-system agentic architectures is simultaneously an existing risk, unanticipated in detection, and emergent over time. It is structural in all deployments where agents coordinate across heterogeneous enterprise platforms, invisible to every current detection methodology because it produces no statistical signal, and accumulates across executions without any individual execution triggering an alert. No current MEASURE 3.1 tracking approach is designed to detect risk that produces clean signals at every individual measurement point while accumulating governance exposure in aggregate.
What the SDI measures and what artifact it produces:
The SDI score at each decision node, recorded in the Semantic Audit Trail, provides the longitudinal record MEASURE 3.1 requires for this risk class. Trend analysis against that record enables detection of gradual definitional drift consistent with Agentic Workflow Drift (2026a).
Sustained SDI divergence patterns may also indicate intentional semantic conditioning, consistent with Agentic Workflow Subversion (2026a), extending SDI from drift detection to adversarial risk identification at the reasoning layer. Neither pattern is detectable through any currently referenced MEASURE 3.1 tracking approach.
MANAGE activation:
Without the SDI, the MANAGE function cannot operationalize intervention for reasoning-layer risk - no signal exists to trigger supersession, escalation, or containment. SDI thresholds define escalation conditions under MANAGE, enabling deterministic intervention based on governed tolerance levels.
MEASURE 1.1 - “Approaches and metrics for measurement of AI risks enumerated during the MAP function are selected for implementation starting with the most significant AI risks. The risks or trustworthiness characteristics that will not – or cannot – be measured are properly documented.”
What the prior submission documented:
The March 2026 submission documented under MEASURE 1.1’s own requirement that
semantic deviation in agentic systems is a trustworthiness characteristic that currently cannot
be measured using available approaches. That documentation was the predicate for the SDI’s
introduction as a candidate construct.
What the SDI paper delivers:
The SDI paper converts the documented cannot-be-measured gap into a proposed measurement approach - the transition MEASURE 1.1 is designed to support. For deployers of agentic systems in regulated financial services environments, this shift is material: the existing requirement to document risks that cannot be measured creates a persistent compliance exposure for reasoning-layer risk. The SDI provides the approach that allows institutions to move that risk from documented-as-unmeasurable to actively monitored, with direct implications for SR 11-7 model risk governance and EU AI Act compliance for high-risk financial services deployments.
MEASURE 2.5 - “The AI system to be deployed is demonstrated to be valid and reliable. Limitations of the generalizability beyond the conditions under which the technology was developed are documented.”
What is being measured:
Semantic validity - whether the agentic system’s definitional resolution at the orchestration layer is aligned with institutional authorization before deployment proceeds. This is distinct from statistical validity and operational reliability, both of which existing frameworks assess. A system can be statistically valid, operationally reliable, and still semantically unauthorized. This establishes semantic validity as a distinct dimension of system correctness, separate from statistical validity and operational reliability.
What the SDI measures and what artifact it produces:
A pre-deployment SDI measurement against the authorized Reasoning Baseline establishes whether the agent’s definitional resolution is within governed tolerance before deployment proceeds. The SDI produces a Semantic Audit Trail and supports Deterministic Gate enforcement, enabling institutions to demonstrate semantic validity prior to execution rather than inferring correctness post hoc. This is the semantic validity evidence that MEASURE 2.5 has no current mechanism to require or collect for agentic systems.
4. What the SDI Paper Delivers
The March submission described the SDI as “a candidate runtime measurement construct requiring industry validation and standardized instrumentation.” The SDI paper (2026b) delivers three components that convert candidate status into a formally defined and computable measurement standard:
Formal specification. The paper defines the SDI formula, its observable inputs (Sₐ and Sₒ), its output range, the conditions under which measurement is taken, and the governance artifacts it produces. It is formally computable, bounded, reproducible, and auditable.
Governance architecture. The paper establishes the three governance components the SDI operationalizes: the Deterministic Gate (threshold-based enforcement halting execution when SDI exceeds governed tolerance); the Semantic Audit Trail (pre-execution governance record of semantic state at each orchestration checkpoint); and the Reasoning Baseline (the authorized definitional reference against which the SDI measures). Together these form the Doyle-Spare Agentic Governance Model (AGM), defined in full in 2026a and 2026b.
Regulatory grounding. The paper establishes the SDI’s relationship to SR 11-7 model risk governance and EU AI Act compliance for high-risk financial services deployments. It is the measurement standard component of a connected research trajectory on record with CAISI across six prior NIST submissions, filed under acknowledgment from NIST/CAISI (March 2026).
5. Request
This submission formally requests incorporation of the SDI paper (2026b) as a reference measurement standard within the AI RMF Playbook under MEASURE 3.2, MEASURE 3.1, MEASURE 1.1, and MEASURE 2.5 in the next update cycle, specifically for agentic AI systems operating in multi-system enterprise environments. The SDI is ready for Playbook-level reference: it is formally specified, governance-architecture-complete, and grounded in a connected body of practitioner research already on record with NIST. No revision to the AI RMF core is required.
The SDI complements, not replaces, existing MEASURE metrics. It addresses the second question that agentic AI deployment requires controls testing to answer: not whether the control fired, but whether it fired against an authorized definition. No existing Playbook MEASURE subcategory currently requires that question to be asked. The SDI is the measurement standard that makes it answerable.
A companion submission addressing Agentic Blast Radius as a containment scoping construct under MANAGE 2.2 is in preparation and will be submitted via this channel when that function’s feedback is active. The full construct is defined in Doyle-Spare (2026b), SSRN Working Paper No. 6531238.
Maureen Doyle-Spare
Independent Practitioner and Researcher, AI Governance and Banking Controls linkedin.com/in/maureendoylespare
INTELLECTUAL PROPERTY: © 2026 Maureen Doyle-Spare. All rights reserved. No part of this article, including its text, figures, tables, and images, may be reproduced or distributed without the author’s prior written permission.
About the Author
Maureen Doyle-Spare
Maureen Doyle-Spare is an enterprise governance architect and independent researcher in AI governance, specializing in agentic AI oversight and cybersecurity for autonomous and multi-agent systems. She is the originator of an agentic AI governance taxonomy built on a set of named constructs, developed for the layer above conventional AI controls, where autonomous agents interpret business meaning across fragmented enterprise systems and commit to it before acting.
Her research develops a runtime governance architecture and a foundational risk and threat taxonomy for the reasoning layer, the pre-execution stage at which an agent’s Operational Interpretation is formed and evaluated. The named constructs include Agentic Workflow Drift and Subversion (AWD/AWS), the Semantic Layer Integrity Attack (SLIA) as a cyber threat class targeting the reasoning-layer attack surface, the Semantic Control Plane (SCP) as a runtime governance mechanism, and the Semantic Deviation Index (SDI) as a measurement standard for semantic divergence. Her central thesis is that agentic systems do not fail the way traditional models fail: an agent can execute flawless steps against a meaning no institution authorized.
Her work spans behavioral visibility, pre-execution oversight, adversarial resilience, autonomous systems evaluation, and safe deployment, and policy crosswalks against the NIST AI RMF, MITRE ATLAS, STRIDE, ISO/IEC 42001, and the EU AI Act. She has provided public comments across multiple NIST AI initiatives, including the AI RMF, CAISI, COSAiS, NCCoE, and TEVV. Her working papers are available on SSRN and Zenodo.
Capstone reference architecture: https://doi.org/10.5281/zenodo.20749051
ORCID: https://orcid.org/0009-0009-6655-1394
SSRN Author Page: https://papers.ssrn.com/Sol3/Cf_Dev/AbsByAuth.cfm?per_id=10836296
ResearchGate: https://www.researchgate.net/profile/Maureen-Doyle-Spare/research
LinkedIn: https://www.linkedin.com/in/maureendoylespare/
GitHub: https://github.com/maureendoylespare/maureendoylespare
Substack:
About this series
Part of the Regulatory Submission Record, a documented archive of practitioner research submitted to standards bodies, financial regulators, insurance regulators and cybersecurity agencies on agentic AI governance, policy and regulation, beginning March 2026.



