The thermometers we didn’t keep
INVESTIGATION · PRE-REGISTERED · DRAFT — FINAL NUMBERS FOLLOW THE ZENODO DEPOSIT
Every argument about the UK temperature record eventually reaches the same two questions. Have the towns crept into the thermometers — is some of the measured warming just concrete arriving around weather stations? And does the Met Office’s processing of the raw numbers push them in a direction of its own? Both questions are old, both are usually answered with rhetoric, and both are answerable with data that sits in the public record.
This investigation answers them the hard way. Every rule — how a station is classed as urbanised or still-rural, what counts as a usable record, what statistical test runs and what any result means — was written down, fought over by an adversarial reviewer until nothing substantive was left standing, and frozen before any temperature was read. The analysis then ran to completion and is reported exactly as it came out. Nothing below was chosen after seeing a result.
What we tested
Hypothesis 1 — urban creep. Station surroundings were classified from four independent evidence streams at every station in the MIDAS network, epoch by epoch from 1975: satellite-measured built-up area, satellite greenness, census population density, and the 1930s Land Utilisation Survey, with airfields screened separately. The thresholds weren’t guessed: they were calibrated on 84 protected nature reserves — places where, by law and verified by zero mapped building footprint, nothing was built — so the gates measure sensor noise, not opinion. A station ejects from the “still rural” pool if instruments that watched empty moorland hold steady say its surroundings changed.
Hypothesis 2 — the processing. Our own homogenisation algorithm — benchmark-proven against synthetic test worlds with known answers, and structurally blind to which stations are urban — ran over the whole usable network. If the official processing style “seeds” rural stations with their urban neighbours’ warming, a blind implementation of the same family of methods will show the mechanism: rural stations’ trends moving toward the towns’.
The first finding is about the record itself
| stage | stations |
|---|---|
| MIDAS stations classified | 1,537 |
| confirmed still-rural since 1975, all evidence agreeing | 46 |
| …of those, with a usable 1975–2018 record (WMO completeness rules) | 6 |
| pre-registered minimum for a viable control group | 30 |
The pre-registered verdict, printed by the pipeline itself: “insufficient confirmable control stations.”
That is not a technicality. A comparison of urbanised against rural stations needs rural stations that (a) provably stayed rural and (b) kept measuring across the whole period. Britain has 46 of the first and six of both. The rural reference network wasn’t destroyed by anyone; it was allowed to close, station by station — of the 46 provably-rural sites, the median record is 14 years and most were shut before 2018. The six survivors sit mostly on high ground, which is why the diagnostic numbers they produce (published in full in the deposit) say more about altitude than about concrete.
The mirror-image number deserves its own sentence: of the 548 stations that do carry a usable 43-year record, 528 — 96% — show measurable built-up growth within 500 m under gates calibrated to nature-reserve stillness. Measurable is not the same as material: the gates trip at a change of roughly two houses’ footprint. But it means the long-record UK network has almost no station whose surroundings are untouched, and any “rural” benchmark chosen less strictly than ours is rural by assertion, not by evidence.
The second finding runs against the sceptical prior
Hypothesis 2 doesn’t need the control group — it tests the processing, not the climate. The blind homogenisation pass made 264 adjustments across the 548-station network. At the six provably-rural stations: zero adjustments. Their trends left the pass exactly as they entered it. The “seeding” mechanism — rural records dragged toward urban neighbours — did not appear. At the urbanised majority the adjustments nudged trends by +0.008 °C/decade on average, beneath our pre-registered materiality threshold of 0.02 °C/decade — and in the opposite direction to removing urban warming, so neither sceptical story survives: the processing is neither manufacturing warming at rural sites nor scrubbing the urban signal.
This lands beside the project’s earlier audit of the Met Office’s own quality-control layer, which found its net effect on mean temperature was −0.019 °C — a slight cooling. Two independent probes of the processing, both pre-registered, both null-to-cooling. On the evidence so far, whatever is in the UK warming record, the paperwork isn’t where it comes from.
For context, on the 71 stations the two share, the international CRUTEM5 dataset’s adjusted trends run 0.05 °C/decade below our raw MIDAS trends — adjustment there also fails to inflate.
What this does and doesn’t say
It doesn’t say urban contamination of the UK record is absent. Hypothesis 1 returned “cannot be confirmed from this evidence”, not “refuted” — the pre-registered design distinguishes those outcomes deliberately, and the deposited power analysis shows the design would have detected a 0.02 °C/decade effect with near-certainty had the control group existed. The question stays open because the evidence to close it was decommissioned.
It does say three things with unusual solidity, because every rule predates every result: Britain cannot currently answer its own urban-contamination question from the raw record; the homogenisation machinery shows no directional thumb on the scale at the thresholds we fixed; and 96% of the stations with long records have measurably urbanising surroundings.
The constructive conclusion is the first one. Six stations is not a policy critique, it’s a maintenance one: the cheapest instrument in climate science is a rural weather station someone keeps running.
Reproduce it
Everything is deterministic from public data. The pre-registration package — specifications, frozen thresholds, calibration data, decision rule, the homogenisation benchmark including its failed runs, and SHA-256 hashes of all of it — is deposited at [DOI to follow]. The pipeline is ukweather.study in the project repository; the analysis re-runs from the MIDAS open archive and the deposited configuration in one command, and verifies its inputs against the deposit’s hashes before it will run at all.
Draft status: the numbers above are the pre-deposit run. The citable run re-executes against the timestamped Zenodo package and will replace them; the frozen hashes make any difference between the two detectable, which is the point of the exercise.