The Anomaly in Sector Twelve
The model is not wrong.
This is the first thing Halima says in every meeting about Sector Twelve, and she says it because it is true and because people keep behaving as though it isn't.
The model predicts traffic flow across nine hundred junctions. It has been in service since 2016. Its median error is under three per cent, which for a system predicting the collective behaviour of two million drivers is close to sorcery.
At junction 12-114 it is out by nineteen per cent.
Every weekday. Not weekends. Between 07:40 and 08:20 the model says one thing and the loop counters say another, and the difference is roughly four hundred vehicles that the model insists should be going a different way.
They have checked the counters. Twice, physically, with a man and a clicker.
They have checked the road geometry, the signal timings, the bus schedule, the school terms, a nearby quarry, and — during a bad week in 2022 — the possibility that a bridge was in the wrong place in the underlying map. It wasn't.
Four years.
Halima has come to think of 12-114 the way you come to think of a difficult relative.
Here is what the model says. Given the network, four hundred drivers should turn left at 12-114 and take the dual carriageway, which is eleven minutes faster on average and has been for the entire period.
Here is what they do. They turn right, into a thirty limit through a residential estate, and rejoin four kilometres later having lost eleven minutes.
They do this every day. The same four hundred, near enough — the plate data confirms consistency, which is the detail that makes the whole thing feel less like noise and more like a decision.
Four hundred people choosing, daily, to be eleven minutes slower.
The theories have been extensive.
Signage: no, they checked, the signage is clear and was improved in 2021 with no effect.
Habit: plausible for some, but the cohort refreshes — people move, change jobs — and the behaviour doesn't decay.
Congestion aversion: the dual carriageway looks busier even when it's faster, and drivers optimise for perceived movement rather than arrival time. This is a real and well-documented effect and Halima believes it accounts for maybe a third of it.
The other two thirds have been a hole in her professional life for four years.
Last month a new graduate called Tobi asked whether anyone had driven it.
There was a pause in the meeting of the kind that is very bad for everybody in the room.
They had not driven it. They had modelled it, counted it, photographed it from a mast, and analysed four years of plate data. Nobody had got in a car at 07:50 and turned right.
Halima drove it on the Tuesday.
The residential route is slower. It is slower in exactly the way the model says.
It goes past a primary school.
Not near it. Past it — along the front of it, at the exact stretch where the twenty limit and the crossing patrol and the parked cars bring you down to about nine miles an hour for four hundred metres, every morning, at eight fifteen, in the middle of the school run.
Halima sat in that queue and watched a woman in the car ahead reach back and hand something to a child in the back seat, and then indicate left into the school gates.
And then the car behind her did the same. And two more, further up.
Not all four hundred. But enough.
The model knows where every road goes. It does not know that a parent will take the slow route because the slow route goes past the school, and that once you are past the school you may as well carry on, and that eleven minutes is not what you are optimising for when there is a child in the back.
Halima went back and reran it with a school-run term in the cost function.
Three per cent.
She has not told the graduate that he solved it, because he did not solve it — he asked whether anyone had driven it, which is not the same thing, and which is much rarer, and which she intends to promote him for.