Ten years of Brexit data, and serious, peer-reviewed studies still disagree by more than a factor of two — from about 4% of GDP to 10%. That spread isn’t noise. Every number answers “compared to what?” — and “what” is a Britain that voted Remain, a place nobody has ever visited.
This is the fourth lab in How Constraints Actually Work. The first three walls each involve someone failing to do something. The Gap Left in the Draft is a wall made of words — a rule with a hole in it. The Duct Tape Doesn’t Care is a wall made of pins — parts nobody stress-tested. The Ruler Stopped Measuring is a wall made of an instrument nobody re-validated. This one is different: the ruler works fine, and the zero mark was placed by the person doing the measuring. Where they put it is the entire argument.
Ten years after the June 23, 2016 referendum, the UK has a decade of post-Brexit data and a mature literature estimating the cost. The estimates are serious, peer-reviewed, and produced by credible institutions — the OBR, Stanford/NBER, NIESR. They also disagree by more than a factor of two, roughly 4% to 10% of GDP. Same underlying real-world data; different answers.
That spread is not sampling noise. Every one of those numbers answers the question “compared to what?” — and “what” is a Britain that voted Remain, a country that does not exist and never will. Analysts construct it: a synthetic control from a weighted basket of comparator economies, a firm-level counterfactual from survey panels, an assumed productivity elasticity. Different constructions, different answers.
When a debate’s confidence intervals are tighter than the gap between studies, the disagreement isn’t statistical. It’s about the baseline. That’s the fourth thing to check before you trust a number: not “is the rule tight,” not “were the parts tested,” not “is the ruler fresh,” but “is the zero mark on this ruler something anybody actually observed?”
● Real — observed, recorded, no counterfactual required. Dates, the referendum margin, prime ministers, sterling’s path, poll numbers.
◐ Constructed — a number computed against a baseline nobody visited. Every GDP-cost figure on this page is this kind. Not wrong — an argument, and one you can only check if you can see its assumptions.
◯ Not yet sourced — a claim this page has not backed with a source. Do not rely on it. (None load-bearing here; every number below is Real or Constructed with a citation on the last tab.)
“I learned this one the hard way. The Auckland incident wasn’t a leak — it was a baseline getting caught. We ran a framework we called international, and it had quietly been measuring the whole world against one Western room. The Pacific was the country that never existed in it. We didn’t punish the person who said so out loud; we fixed the problem the mistake revealed. Do the same with any number they hand you: don’t start by asking if it’s big. Ask what room it was measured in — and who wasn’t in the room.”
| The country that never existed | |
|---|---|
| Parameter | What actually happened to the UK — observed, recorded, not in dispute. Prime ministers, the referendum margin, sterling’s path, poll numbers. |
| Variable | The constructed counterfactual — the Remain-Britain the real one is measured against. Built, not observed. It moves with methodology. |
| Resource | Attribution budget. Covid, the energy shock, global trade fragmentation and Brexit all happened at once. Every point of damage charged to one is a point not charged to another, and the total is finite. |
Drag the two sliders. The headline number — “Brexit cost the UK ___% of GDP” — moves across roughly the 2%–11% range as you drag. The real data never changes. Only the baseline does. ◐ the plane is mine — a teaching synthesis, not a measurement.
The vertical axis is a finite pool, not a dial. Confounder attribution isn’t just “how much to blame Brexit” — it’s Brexit’s slice of a fixed total it shares with Covid, the energy shock, and global fragmentation. Raise it and you are taking those points from the other causes, not inventing new damage.
The trap corner (top-right): high generosity × high attribution produces the largest number, and it looks exactly as rigorous as the smallest one. Nothing in the output signals that two assumptions were doing the work — and, deliberately, there is no confidence band on this display to narrow or widen. The stated confidence would be blind to the baseline choice anyway.
The instructor hands you a policy decision from your own life or work — a job change, a road project built or not built, a program funded or cut. Construct the counterfactual. State explicitly what you’re assuming about the world where the other choice was made. Then defend, in one paragraph, why your baseline isn’t just the answer you wanted, written backward. There is no right coordinate. There is only a baseline you can argue for.
This is the reason to build the lab. The dates and the poll were observed. The percentages were computed against a Britain nobody has ever visited. Both belong on the page — but the page says which is which.
The three earlier walls all involve someone failing to do something — draft a rule tightly, test an assumption, re-check an instrument. This wall involves everyone doing their job correctly. The synthetic-control method is legitimate. The decision-maker’s panel is legitimate. The disagreement survives full competence and full good faith — because the object of measurement, the road not taken, is not observable in principle. A published 500-year discharge and a published Brexit GDP cost are the same species of artifact: a confident number whose confidence lives in the method, not the observation.
Why it matters. The loudest number in a policy fight is usually the one with the most generous hidden baseline. This lab is neutral on whether Brexit was good — the 10% read and the “smaller than feared, fading” read get equal, fair placement on the plane. The point is the epistemics: a number you can’t see the assumptions of is a number you can’t check. If a reader can tell how the builder voted, the lab is built wrong.
Referendum date and margin, the run of prime ministers, sterling’s path, the June 2026 poll, ONS demography, and the pace of retained-EU-law reform are recorded facts, sourced below. Each is quoted to its source, not to a counterfactual.
The confidence-vs-baseline framing and the two-axis plane are mine — a teaching synthesis, not a measurement. So is the “fourth wall in the constraint set” framing and the sibling closer. The plane’s headline number is an illustrative interpolation across the real published range (~2–11%), not a model output; its presets are pinned to real studies but its intermediate values are a teaching device. There is deliberately no confidence band, because a stated confidence would be blind to the baseline choice — which is the whole lesson.
This page carries citations, but the numbers split into two kinds and the page says which is which. The dates, the margin, the prime ministers, the poll — those were observed. The percentages were computed against a Britain nobody has ever visited. That doesn’t make them wrong. It makes them arguments, and an argument you can’t see the assumptions of is an argument you can’t check. (Illustration: a 10% figure that assumes a Remain-Britain tracking EU27 growth and charges nearly all of the shortfall to Brexit is making two choices, not measuring one effect — a page should say so, not let the number imply a pure Brexit reading.) Spot something off? Email User Zero — corrections get acknowledged right here.