UKRI Future Leaders Fellowship: Nominated, Round 11
Selected as one of two nominees at Leeds Beckett Law School, from a pool including doctoral candidates. June 2026.
Patsy Nwogu
Computational Governance · AI Policy
Patsy Nwogu is a computational governance researcher and practitioner. Across seven years in data analysis, risk management and AI governance, in public and private sector settings, she has moved from analysing risk to building the systems that govern it, most recently as AI Governance Lead at City Business AI (Traavu Ltd), where she designs the guardrails that keep deployed AI systems compliant and operationally sound. She also works directly in specialist SEND education, supporting Year 7 to 10 learners with moderate to severe autism, work that grounds her research into safeguarding governance in direct practice rather than theory.
Her research rests on one principle: policy should be tested against the conditions it will actually operate under, before it is deployed rather than after. Computational governance encodes a policy's operating assumptions as parameters and tests them against realistic conditions, using whatever quantitative method the question calls for, simulation, statistical modelling, or optimisation, so the gap between design intent and operational reality becomes measurable rather than assumed. This is not a methodological preference. It is a form of risk management, one that surfaces failure before it reaches the people a policy is meant to protect.
A distinguishing feature of the approach is the pairing of a rigorous computational backend with a stakeholder-facing implementation. A model that only researchers can read cannot be improved by the people it represents. When practitioners can adjust the operating conditions themselves and watch the consequences in real time, they become participants in a model's validation rather than subjects of its conclusions.
The Method
Policy is usually assessed after it has been implemented, when the record of what went wrong is the only evidence available. Computational governance treats it as testable in advance.
A policy framework carries assumptions about the conditions it will operate in, that staff have time to record what they see, that an algorithm's outputs will be reviewed by someone able to contest them, that the infrastructure a regulation depends on is actually present. Where those conditions don't hold, the instrument can be sound and still fail. The method encodes those assumptions as parameters in a computational model, agent-based, statistical, or optimisation-based, depending on the question, and tests what they produce before the policy meets the people it governs. Policy enters the analysis as the parameter regime itself, not as an external input.
What this produces is not a prediction about individuals. It is a stress test of an architecture: a statement of which operating conditions a policy depends on, and what happens when they are absent. Any framework whose success rests on such conditions can be audited the same way, which is what makes this a method rather than a single study.
Research
Research output in computational governance, spanning safeguarding, AI governance and financial regulation.
The simulation, running
A walkthrough of the GRID Observatory: agents interacting under the baseline, naive AI, and governance conditions described in the paper, with the recording floor and escalation routing visible as the week plays out.
Run simulation →GRID Observatory: simulation walkthrough
The Mathematics of Safeguarding
Safeguarding Policy · Agent-Based Simulation · SEND Education
Non-verbal learners in specialist SEND schools cannot disclose harm, so safeguarding depends entirely on what practitioners record. This paper defines the gap between what staff observe and what reaches the institutional record as underreporting, gives it a testable form, and encodes current DfE guidance as the baseline parameter regime of an agent-based simulation. Across 100 Monte Carlo replications the baseline produced a crisis in every replication; a naive AI layer accelerated the failure; a governance layer eliminated crises and raised classified records 4.4-fold.
One Instance of the Method
In this application the operating conditions reduce to three that must hold at once for a safeguarding event to reach the record. Because they are conjunctive rather than compensatory, a value near zero on any one drives capture toward zero whatever the other two are doing.
U = 1 − (C × A × T)
U is the probability that a complete event fails to reach the record.
Underreporting
0.969
96.9% of events lost
Capture
0.031
C × A × T
Marginal return
C
Raising C returns most here
Sensitivity: What Each Condition Buys at the Current Position
Holding the other two fixed, underreporting is linear in each condition. The slope of each line is its marginal return, and every slope is the product of the other two. Move a slider and watch the other lines change gradient.
varying C
varying A
varying T
This is the closed-form expression, not the simulation. The agent-based implementation, Monte Carlo results and sensitivity sweep are on GitHub. Baseline values are those used in the paper: C = 0.25, A = 0.50, T = 0.25. The formalism itself is not specific to education. It applies wherever a population that cannot self-report depends on staff observation reaching a record.
The Environmental Gap in Agentic AI Governance
AI Policy · Infrastructure Assessment · Pre-Deployment Governance
The EU AI Act, the NIST AI Risk Management Framework and the African Union's Continental AI Strategy all treat human oversight as primary to safe deployment. This paper does not dispute the principle. It questions the assumption that the conditions enabling oversight are stable and reliably present. Connectivity drops, institutions are under-resourced, operators are fatigued or undertrained: the human in the loop is only as effective as the conditions around them. The argument draws on a deployment incident the author led involving an agentic voice system in an underground venue on an unstable connection, on published evidence about infrastructure conditions in the Global South, and on Dagstuhl Seminar 25272, which identified environmental factors as one of three conditions for effective human oversight.
Simulating African Fintech Futures
Financial Policy · Simulation · Built Environment
Financial regulation in emerging economies is written for conditions that often do not hold in the settings where it operates. This work simulates fintech policy within its built environment, testing regulatory design against the infrastructure it actually depends on.
Recognition
Think Big Scholarship
Awarded 2026.
Research Advisory Board
Three professors in agent-based modelling, geographic data science and complexity economics, each of whom engaged with the research independently before agreeing to advise. Names available on request.
AI Governance Lead, City Business AI (Traavu Ltd)
Applied research on AI-driven decision-making and data governance in operational settings.
Mentorship
Mentored by a UKRI Future Leaders Fellow at Leeds Beckett Law School.
Data Analyst (Assistant Lead), Orodata Science
Led the replacement of unverifiable COVID-era health data across 345 Primary Health Centres serving 1.6 million beneficiaries, imposing verification thresholds before each tranche release. Rebuilding that risk case visually, after statistical analysis failed to land with non-technical stakeholders, is where the stakeholder-facing design principle in this method began. 2021–2023.
Open Simulation Code
The GRID Observatory implementation, figure-generation scripts, sensitivity sweep, Monte Carlo summaries and per-replication results are public, with a deterministic seeding scheme so any reported run can be reproduced.
Journaling My Thoughts and Research
Notes on computational governance and AI policy, written for anyone rather than for a journal.
Contact
For research collaboration, policy work, or questions about the simulation.