By Patrick Lloyd
In a statement issued on 22 October and updated on the 26th, the Australian Bureau of Meteorology answered Tom Berger’s audit of Australia’s raw temperature record. It confirmed “a small number of duplicated historical daily temperatures” in four post-1910 months relevant to ACORN-SAT, its adjusted national series: “fewer than 200 daily values”, which it said it had corrected. It added:
“Further investigation has detected no further duplicate months in the entire bureau network other than those quoted in the report and then confirmed by us.”
I keep a full copy of the Bureau’s public daily data: every station’s raw maximum, minimum and rainfall, downloaded from Climate Data Online. I took one copy in May and another on 29 and 30 October. That let me check two things: whether there are other copied months, and exactly what the Bureau changed.
There are more copies
I searched every station’s raw daily maximum and minimum for stretches that repeat another stretch of the same record, day for day, exactly or (in one case) to within rounding. Then I checked each one against neighbouring stations to see which block was the original and which was the copy.
The result is 44 copied stretches at 15 stations, 978 days in all. They run from 1861 to 1981. Thirty-two of them, 640 days, are still in the files the Bureau publishes today. Some examples, all downloaded from the Bureau on 29 October:
- Cairns: December 1950 maximums are a copy of January 1950, and the minimums match too.
- Toowoomba: February 1981 maximums repeat February 1982.
- Charleville: August 1914 maximums repeat September 1914.
- Weeaproinah (Victoria): February 1975 maximums repeat February 1974.
- Inverell: April to June 1875 and the same weeks of 1878 are identical.
- Bathurst Gaol: January 1862 minimums repeat December 1861.
- Sydney Observatory Hill: 13 days in October 1890 are identical to the same days in November 1890, maximum and minimum.
None of these is in Tom Berger’s audit, so the Bureau’s exception for duplicates “quoted in the report” does not cover them. “No further duplicate months in the entire bureau network” is not true of the data the Bureau publishes. It may be true of ACORN-SAT, the adjusted record, which takes in only part of the raw record. If that is what the Bureau meant, it should say so.
What the Bureau actually changed
Between my May and October downloads, 593 temperature values disappeared from the raw files. Twenty-two of them are routine revisions to 2025 and 2026 data. The other 571 are at just two stations: Sydney Observatory Hill (515) and Charleville (56). No other value before 2000 was removed. One was added: a maximum of 0.0°C for Sydney on 25 June 1865, flagged as quality-checked, in a month the Bureau had just emptied. The copies at Cairns, Toowoomba and the rest are still there, and so are the ones at the two stations it fixed: Sydney’s October 1890 and Charleville’s August 1914.
Where the Bureau deleted one month of a copied pair, it deleted the right one: the copy, by my neighbour test. But for seven of the eight Sydney pairs before 1910 (the eighth, October/November 1890, is untouched), it deleted both months, the copy and the original. If one month in each pair is genuine, as one would expect, roughly 200 genuine readings went with the copies. For two of them my neighbour test can tell which month is genuine: May 1883 and November 1905. Both are gone.
The Bureau also says it “does not overwrite or delete original temperature observations when producing ACORN-SAT”. That may be so, but the published raw record has now had 571 values removed, and the statement doesn’t list them. The only way to find out was to compare downloads, as I did.
A seven-day match still inside ACORN-SAT
Sydney’s minimums for 25 to 31 January 1910 are identical to those for 1 to 7 February 1910, seven values to a tenth of a degree, with the days either side different. Tom Berger’s audit identified it. It is not among the 44 above because seven days is too short to clear the statistical bar I set for a national search, but it has every appearance of a copy. That week is still in the raw file, and ACORN-SAT keeps both weeks. ACORN-SAT adjusts January and February by different amounts, so in the adjusted record the two no longer match exactly. Apart from the last five days of Charleville’s copied August 1914, where ACORN-SAT’s record resumes after a gap, it is the only such match I have found inside ACORN-SAT. Seven days at one station make no measurable difference to any trend. If it is a copy, the Bureau’s check of ACORN-SAT missed it.
The rainfall is worse
The Bureau’s statement was about temperature. So I ran the same kind of search on every daily rainfall record on Climate Data Online, all 17,898 stations, looking for stretches that repeat another stretch of the same station’s record exactly, to the 0.1 mm. To make sure I wasn’t counting chance matches, I ran every search again on control data built to contain no real copies (the same series shifted by a few days, or read backwards). Nothing in those controls passed the bar.
The result: 1,337 copied stretches at 1,191 stations, in every state. Together they cover 61,877 days, 13,964 of them with rain. The rain days alone are about 70 times the “fewer than 200 daily values” the Bureau reported for temperature. They run from the 1860s to this year, peaking in the 1970s and 1980s, and 60 are from 2000 or later. The Bureau put its temperature copies down to hand-digitised records from before 1946. Copies from this century can’t be explained that way.
Some of them are long. At Muresk in WA, 381 of the 402 days from January 2007 to February 2008 are identical to the corresponding dates six years later. At Iawakia, also WA, most of 1984 matches 1981. At Bagdad in Tasmania, August to December 1903 matches the same months of 1904.
The copies are marked as checked. More than 97% of the copied stretches carry nothing but the Bureau’s “Y” (quality-checked) flag, the same rate as genuine data. Its quality control didn’t catch them.
Repeated rainfall across batches of stations
One group stands out. At 22 stations in six states, rain from 1998, 1999 and 2000 has been written into 1980, 1990 and 1997: onto the same dates for the first two batches, and onto February to April 1997 for September and October 2000. Several stations share one exact offset in each batch, across state lines. That is not an observer copying the wrong page. It looks like a database operation.
Checking against neighbouring stations shows which block is real. Wherever it can be placed, it is the late-1990s rain that fits its own dates’ weather, and the earlier years that don’t. At Pokolbin in the Hunter, September 1980 now holds September 1998’s rain, including 24.2 mm on the 6th, when four nearby stations average about 0 mm. The earlier years sit inside otherwise continuous records. Whether genuine readings were overwritten or gaps were filled with another year’s rain, the public file no longer holds what fell on those days.
Other stations’ rain
Rainfall also turns up under the wrong station. 1,649 pairs of stations carry identical stretches on the same dates. Most are one gauge published under two numbers, or a return filed under a station number one digit off. But 47 pairs at least 50 km apart have no explanation I can find. Waterloo in Queensland and Krambach in NSW, 821 km apart, are identical for 358 of the 366 days of 1964. The rain fits Waterloo’s district, so Krambach’s 1964 record is Waterloo’s rain.
1,334 of 1,337 unchanged
I downloaded every rainfall station again on 30 October (17,896 of the 17,898 came through) and compared each, day by day, with my May copy. 1,334 of the 1,337 copies are still published unchanged. The other three are all recent and look like routine revisions. Nothing at all before 2000 has changed. The Waterloo–Krambach pair and the overwritten batches are still there.
Does it matter?
The copies touch a small fraction of the archive: by a neighbour-based estimate they carry a net excess of about 4,180 mm, against some 59 million mm held in the records of the 1,191 stations concerned. I have not quantified their effect on national rainfall trends. They don’t target droughts or floods, and they run backwards in time as often as forwards. Like the temperature copies, they look like data handling errors, not anyone’s agenda.
But for anyone who uses a single station’s record, it matters a great deal: farmers, flood and drought studies, insurers, water allocation. At those 22 stations, the 1980, 1990 and 1997 rainfall is another year’s rain, and the Bureau’s own flags say it has been checked.
What this is, and what it isn’t
To be fair to the Bureau: these look like clerical and data handling errors, not tampering. The temperature copies show no measurable lean warm or cool (15 warmer, 10 cooler, where the direction can be told). None is more recent than 1981, and ACORN-SAT, the series that carries Australia’s warming trend, excludes all of them bar the Sydney week and five Charleville days. The rainfall copies are a small fraction of the archive, and I have not quantified any effect on national trends.
But the Bureau’s statement is wrong on the facts:
- There are further duplicate months, and most of them are still published.
- “Fewer than 200 daily values” describes only the four post-1910 months relevant to ACORN-SAT. Correcting two stations removed 571 values from the published raw data, most of them pre-1910.
- In fixing Sydney it deleted genuine nineteenth- and early twentieth-century readings, without saying which ones.
- It fixed two temperature stations and left the rest. In rainfall, 1,334 of 1,337 copies are still published.
The raw record is the evidence everything else rests on. If errors are found in it, the Bureau should fix them all, in temperature and rainfall, publish a list of what it changed, and restore the genuine readings from the paper ledgers rather than deleting them.
Run it like an open-source project
There is a better way to look after a national record, and the software world worked it out decades ago. Open-source projects publish everything: the code, every change ever made to it, who made it and why. Anyone can find a bug, report it, and propose a fix. Fixes are reviewed in public, and nothing is quietly deleted.
Australia’s weather record could work the same way:
- Scan and publish the originals. Every observation book, ledger and register the Bureau holds, as images anyone can download, so any digitised value can be checked against what the observer actually wrote.
- Keep a public change log. Every value added, altered or removed, with the date and the reason. Right now the only way to find out what changed is to download the whole record twice and compare it, as I did.
- Take corrections from anyone. A farmer with a family rain book, a historian with a newspaper weather column, a retiree with a spreadsheet. Give them a way to flag a problem, show their evidence, and see it resolved in the open.
- Never delete, supersede. When a value turns out to be wrong, mark it and replace it, but keep the old one visible with a note. A record that can be silently edited can’t be trusted, even when the edits are right.
None of this is new. Scanning the ledgers is real work, but the tools for the rest already exist: Wikipedia, OpenStreetMap and Linux are all maintained this way, by thousands of volunteers, and they are better for it. Every copy described here was found by people working from home, using data anyone can download, and the Bureau’s own quality flags mark almost all of the rainfall copies as checked. Give them the original records, and they will help the Bureau reconcile and repair the rest.
I’m happy to share the full lists of copies, temperature and rainfall, with dates and stations, with the Bureau or anyone else who wants to check it.
Thanks to Tom Berger and Geoff Sherrington, whose audit started this.
Sources: the Bureau’s statement, bom.gov.au/news-and-media/the-science-and-truth-of-australias-long-term-temperature-record; Tom Berger’s audit, elastictruth.com/2026/07/a-forensic-analysis-of-systematic-data.html; the raw data, Climate Data Online, bom.gov.au/climate/data.
