Putting the AI in ApplIed Urban Modelling
A Complete Model of Brighton and Hove’s Secondary School System in Five Months
Friday 16 October 2026
Who I am, and why this talk
- Professor of Urban Analytics at UCL’s Centre for Advanced Spatial Analysis. I build spatial models — of migration, of housing, of how people move around cities.
- Also a parent in Brighton & Hove, and one of hundreds of families drawn into a two-year public debate about secondary school admissions.
- I spent two years analysing the drivers of secondary attainment in England, published as How to Pull the Right Lever.
What I found: the council’s policies were unlikely to improve attainment; the costs — longer journeys, community schooling breaking up — were not weighed against them; and absence, the single largest factor, was not part of the decision at all.
A misalignment runs through the whole episode, between the council’s stated aims (improving disadvantaged attainment), its stated problems (falling rolls, school viability) and the policies actually chosen (admission priorities, catchment boundaries, PAN cuts) — with little attention paid to how those interact, or to what they cost elsewhere in the system.
The cohort is already falling
Income is tied to numbers - and financial reserves are running out
School Allocations and Viability
The city placed 1,900 children in 2010 and 2,250 in 2026. Over the same years Longhill fell from 239 to 81, and Hove Park from 308 to 136. Now the pupil roll is declining, serious problems ahead
An AI4CI problem, if ever there was one
The UKRI National AI Research Hub for Collective Intelligence - 5-year, £12m EPSRC Project - asks what AI can do to enhance the thinking groups of people do together and the solutions they can arrive at.
Can AI help us develop a shared, inspectable account of how the school system in Brighton and Hove actually behaves?
The challenge: If AI can help us get to an answer quicker - can I ask the right questions? Can this collaborative effort enhance our Collective Intelligence?
A system where everyone is well informed about a part and nobody can see the whole is not an information problem. It is a collective intelligence problem.
What the model had to learn
Three things are fitted to the council’s own published allocations, rather than assumed:
Distance decay, \(\beta = 1.7\). How steeply demand falls as a school gets further away.
The catchment pull, \(\gamma_h\). How much living inside a catchment raises the odds of choosing a school there — fitted separately for each catchment, because they do not behave alike.
Competition, \(\delta = 0.3\). Whether being surrounded by rivals helps a school or hurts it.
Living in the catchment multiplies the odds by 2.7× in Brighton Aldridge’s and 13.5× in Stringer and Varndean’s — one policy doing five times as much work in one place as another.
Fitted this way, the model reproduces the 2026 offers with an R² of 0.95.
Attractiveness - \(W_{j}\) = Exam Scores
The council’s secondary admissions guide tells parents to look at each school’s “National Curriculum test results, public exam results” and its Ofsted reports: figures that mostly describe who a school admits.
A test the model could fail
The disadvantage layer is fitted to each school’s published disadvantaged share, so reproducing that share proves nothing. The FOI arrived weeks later with something the model had never seen: how many children each school actually admitted under each priority.
In catchment it predicts well — R² = 0.94, average error 4.1 children.
Out of catchment the total is right and the destinations are not: 95 children modelled against 99 actual, but at the wrong schools. One does nearly all of it — the model places 34 at Varndean; 6 went.
Why Varndean, and why no parameter fixes it. The council admits siblings under priority 3, above both free-school-meal tiers — 70 children at Varndean. The model has no priority 3: looked-after children, SEN and siblings cannot be told apart in zone-level flows.
So a disadvantaged child with a sibling there lands in the model’s tier 5 and in the council’s priority 3. The two numbers are not counting the same children.
I tried the obvious fixes — letting these children decay with distance on their own terms, then flattening how much a school’s reputation pulls them. Steeper decay makes the fit worse on every measure: the city-wide total was already right, so suppressing travel breaks a number that worked. Flattening attractiveness helps slightly and leaves Varndean at 31 against 5. It is a definitional gap, not a behavioural one — which is a more useful thing to know, and took an FOI, a sweep that failed, and the project’s own model guide to establish.
Worked example: the simulator
- Simulator has both pre-built scenarios I imagined plus scope to change each element.
- Built in about two days. “what if” questions became cheap to ask multiple times
Back to the first question
“What is causing Longhill’s acute numbers problem — is it geography, demography, attractiveness, some combination or something else? Would moving and/or shrinking it help it survive, or improve anything else?”
- The honest answer is none of them on their own.
- Longhill is not short of children for one reason, so no single lever fixes it.
- What the simulator lets you do is pull them one at a time, and watch them compound.
- Today, 2026.
- Longhill takes 90 of 210 places (modelled) (43%), 120 short.
- A disadvantaged child travels +3.4 minutes further than everyone else.
1st Lever: families read value added
One dropdown. Families read the four-year value added measure — what a school adds once intake is accounted for — as closely as they now read the headline score. Nothing about the schools changes.
Longhill goes from 90 to 106 of 210 (50%).
The travel gap falls from +3.4 to +2.5 minutes, the city’s mean journey drops by 42 seconds, and segregation rises slightly.
2nd Lever: move it, restore its catchment
Two more controls: Longhill relocates to Elm Grove, and Whitehawk returns to the Longhill catchment it sat in before 2024.
Longhill reaches 140 of 210 (67%), and the travel gap reverses: disadvantaged children now travel 1.9 minutes less than their classmates. Segregation rises, but no child is displaced from a central catchment.
Redrawing the catchments as flow regions around the new site was also tested. It did worse — a travel gap of −1.0 against −1.9, and twice the displacement.
3rd Lever: and if it taught like BACA
The last lever is not an admissions policy. It is the school getting better — value added doubled, to roughly where Brighton Aldridge (BACA) stood over the four years to 2024/25, on a comparably disadvantaged intake.
Longhill reaches 197 of 210 — 94% full, 13 places short, 1% of children displaced city-wide.
A disadvantaged child now travels 2.2 minutes less than everyone else — a swing of 5.6 minutes. Absence is this city’s largest attainment problem, and distance is one of the few things policy can change about it.
Why this needs a simulator
Each choice is modest on its own. 7 points, then 17, then 27. None would survive a committee on its own merits. Together they take a school from 43% full to 94%.
Segregation rises, 0.117 to 0.188. That is the honest cost.
But How to Pull the Right Lever finds the attainment model disagrees about direction between disadvantaged and other pupils — so deconcentration bought with longer journeys is unlikely to help the children it is meant to.
You cannot reason about this in a committee paper. Every lever moves every other one, and you have to see each choice land.
Five months, in one chart
Counted, not remembered: ~250 MB of Claude Code session logs across ten sessions, plus the git history of six repositories. 1,277 instructions from me, 13,451 tool calls back, over 69 working days spread across 166 — about ten and a half actions for every sentence I typed. A snapshot; the work is still going.
Figure 3
The April–July work is a different project — a national school attainment model. The Brighton system model starts on 8 September and is finished eighteen days later.
