Home About Research People Papers Contact

Computational
Meaning
Dynamics

We study how meaning moves -- measuring, modeling, and exploiting semantic instability in AI systems to advance both security and scientific understanding.

Scroll

Recent Activities

Aug 2026
Advancing AI Governance in Healthcare
TwinPath AI is an AI governance architecture that ties multi-axis semantic evaluation to longitudinal care pathways, so AI support can be measured against actual patient-care trajectories rather than judged in isolation. Built on Computational Meaning Dynamics methods, it exposes misalignments in both AI outputs and care behavior retrospectively and as it emerges, giving clinical and quality teams a data-grounded way to monitor, audit, and govern AI and care together. Learn more about TwinPath AI.
Jun 2026
Visualizing Meaning In AI
Meridian Observatory is a live-computing research instrument that measures the geometry of meaning as a conversation unfolds. Built on Semantic Substrate Dynamics Theory, it treats each interaction as a trajectory through embedding space and computes how meaning moves, turn by turn, in real time. Its rare capability is generating clean, labeled, per-turn-embedded interaction data under controlled conditions, then using it to track how meaning holds, drifts, or collapses across AI-to-AI and human-AI exchanges. This makes semantic dynamics and manifolds observable and measurable as they emerge rather than after the damage is done; read about Meridian Observatory.
May 2026
Semantic Sensemaking Surveillance
SemPythia was created as a research workbench that measures and visualizes semantic frame shifts across live global event-streams. It ingests real-time open-source news and events as a layer on an interactive world map and watches how events are situated and get contextually reframed over time. For each shift it computes the geometry of meaning directly: how events’ details move through latent embedding space, how their semantic neighborhood reshapes, and how the distribution and entropy of meaning shift across temporal windows, turning reframing into a labeled, geo-spatial drift trajectory. Follow the work.
Dec 2025
Measuring Semantic Drift in AI Knowledge
DARMA (Drift-Aware Retrieval Metrics and Alignment) is a research instrument that measures whether a revision actually changed what a document means. Retrieval-augmented generation grounds AI answers in documents that are revised constantly, proposals, guidelines, program specs. Seeing that text changed is easy; knowing whether the change moved meaning, and by enough to shift an answer, is not, and across a whole knowledge base is even harder. DARMA treats every document chunk as a point in embedding space and reads the geometry of how meaning shifts between versions, separating substantive revision from surface rewording. Every revision gets a calibrated signal: which changes matter, where, and how much. This makes semantic change in a knowledge base observable and measurable as it happens, rather than discovered after it has shaped a decision. Learn how DARMA works.
Sep 2025
Personalized Meaning Matters
AiVisor is an agentic advising system that treats personalization as an experimental variable rather than assuming it helps, measuring what it actually does to answer quality. Built on Computational Meaning Dynamics methods, it runs personalized and non-personalized configurations across lexical, semantic, and reasoning metrics under deliberately conservative conditions. The AiVisor research isolates a trade-off standard evaluation hides: personalization improves reasoning and grounding while standard semantic metrics penalize it for correctly deviating from a standard reference, exposing how reference-based similarity metrics misjudge personalized answers. See an infographic about AiVisor.

Research Areas

Core

Semantic Drift Detection

Active
SSDTSemantic DriftRecursive LLM GenerationDynamical Systems

Semantic Substrate Dynamics Theory (SSDT) is the foundation the group is built on: the claim that meaning behaves as a dynamical system with measurable structure. Word senses, diagnostic codes, and learned representations do not hold still. They shift when models process their own outputs and when real-world usage moves underneath them. SSDT gives us the vocabulary and the geometry to measure that motion, track where it accumulates, and mark the point where it crosses from noise into failure. See DriftWell as an example

Methods

Geometric Methods in AI Analysis

Active
Ollivier-Ricci CurvatureEmbedding GeometryFrame-Shift Detection

Meaning has shape, and shape can be measured. This area applies Ollivier-Ricci curvature and related tools from differential geometry to the embedding spaces where language models hold meaning. Curvature marks where a semantic region is tightly bound and where it is about to fragment. Graph structure shows how concepts connect and where those connections reorganize. Followed over time, the same geometry reveals where meaning bends, where a frame shifts, and where a trajectory is about to break. It turns semantic movement into something with coordinates, direction, and rate. Read about Ollivier-Ricci Curvature in SSDT

Applied

AI Reliability

Active
RAGAgentic PipelinesEvaluationAdversarial Robustness

Agentic systems chain retrieval, reasoning, and generation, and meaning can move at every hop while the workflow looks stable. We study where that happens and what it costs. A pipeline can carry the right answer forward or quietly reframe it, and the difference rarely shows on the surface. The same instability that degrades reliability by accident can be induced on purpose, which is why adversarial semantic manipulation sits inside this work rather than beside it. The question throughout is whether an agentic pipeline still means what it started with by the time it answers. See how AiVisor exposed a structural flaw in how these systems get scored

Exemplar

Clinical AI and Healthcare Informatics

Active
Coding DriftRevenue CycleClinical SurveillanceAI Governance

Healthcare is where semantic drift stops being academic. Clinical ontologies shifting or a governance model that loses clinical specificity causes changes in meaning-in-use that carry direct costs and patient-care risk. Our work focuses on coding drift detection and shifts across large inpatient cohorts, denial and revenue cycle prediction, VTE and SSI surveillance, and governance frameworks that keep AI aligned to care trajectories. See AI governance in TwinPath AI|Why governance is needed

Frontier

Human-System Integration and Agent Orchestration

Active
Human-Agent TeamingOrchestrationSituation AwarenessGovernance

As agents take over ingestion, triage, coding, and other pipeline processes, humans are increasingly disconnected from direct evidence and start seeing workflow products that upstream AI has already shaped. An object can look stable while the meaning underneath it moves -- from accident to attack, or from outage to state-linked operation -- with nothing on screen to signal the shift. This area builds the supervisory layer that catches it: semantic-dynamics metrics that flag when a stable object no longer carries stable meaning, and orchestration that routes reframing and human-agent disagreement to a person while letting settled outputs pass. The goal is to keep the human in command of meaning, not just in the loop. See an example analyst tool with semantic frame-shift capabilities

People

Stephen Russell
Stephen Russell
Professor, Intelligent Systems and Robotics Department
Semantic Drift, Human-Machine Teaming, Healthcare AI, Decision Support Systems, Intelligent Agents
Graduate Researchers
Maria Ramirez
Maria Ramirez
PhD Student
Semantic Instability, Data Science
Satyajit Movidi
Satyajit Movidi
Graduate Research Assistant
Semantic Equivalency, RAG Systems, Personalization
Pratigya Paudel
Pratigya Paudel
Graduate Research Assistant
Uncertainty Quantification, Machine Learning
Heather Anderson
Heather Anderson
Volunteer Research Assistant
Healthcare AI, Machine Learning

Select Papers

2026

The Personalization Paradox: Semantic Loss vs. Reasoning Gains in Agentic AI Q&A

Cloud Computing and Data Science
Satyajit Movidi, Stephen Russell

AIVisor, an agentic retrieval-augmented LLM for student advising, was used to examine how personalization affects system performance across multiple evaluation dimensions. Results showed a consistent trade-off: personalization reliably improved reasoning quality and grounding,...

2026

When Workflows Stay Stable but Meaning Moves in Agentic Analyst Pipelines

Human Factors in Robots, Drones and Unmanned Systems, AHFE International
Stephen Russell
DOI
2025

VTE Surveillance: AI Enabled Human Machine Teaming for Veinous Thromboembolism

Stephen Russell, Fabio Montes Suros, Venessa Goodnow, Jia Zeng
2024

Exploiting Machine Learning Bias: Predicting Medical Denials

Proceedings of the Association for the Advancement of Artificial Intelligence Spring Symposium 2024
Stephen Russell, Fabio Montes Suros, Ashwin Kumar

For a large healthcare system, ignoring costs associated with managing the patient encounter denial process (staffing, con- tracts, etc.), total denial-related amounts can be more than $1B annually in gross charges. Being able to predict a denial before it occurs has the...

DOI
2022

Providing Care: Intrinsic Human-Machine Teams and Data

Entropy
Stephen Russell, Ashwin Kumar

Despite the many successes of artificial intelligence in healthcare applications where human–machine teaming is an intrinsic characteristic of the environment, there is little work that proposes methods for adapting quantitative health data-features with human expertise insights....

2021

Re-orienting towards the Science of the Artificial: Engineering AI Systems

Systems Engineering and Artificial Intelligence
Stephen Russell, Brian Jalaian, Ira S. Moskowitz
2021

Towards Safe Decision-Making via Uncertainty Quantification in Machine Learning

Systems Engineering and Artificial Intelligence
Adam Cobb, Brian Jalaian, Nathaniel Bastian, Stephen Russell
2020

Separating the Forest From The Trees - Wavelet Contextual Conditioning For AI

Human-machine Shared Contexts
Stephen Russell, Ira S. Moskowitz, Brian Jalaian
URL
05

Contact

We welcome inquiries from prospective graduate students, collaborators, and organizations interested in supporting this research. If your interests involve meaning and reliability in AI, human-AI systems, or applied AI research, we would be glad to hear from you.

Dept. of Intelligent Systems and Robotics, University of West Florida
Pensacola, Florida