Abstract

In October 2026 The Australian published a series of articles built on a 40 page audit of the Bureau of Meteorology's raw temperature archive, written by Tom Berger with data work by Geoff Sherrington. The Bureau answered on 22 October with a statement calling the manipulation claims "false and unsubstantiated" while conceding four duplicated months after 1910. The Guardian's Graham Readfearn followed on 25 October asking where the evidence was, in a column that reports no check of the data itself. We went and looked. The duplicated months are real, and they are still sitting in the file the Bureau publishes as "raw" in the very same October 2026 bundle in which it says they were fixed. The Palmerville flatline is real but has a mundane explanation the audit should have found. The claim that every year of Sydney's maximum temperature was adjusted upward is real, we count 107 of 107 years to October 2017. None of the three parties to this argument looked at the question that actually matters, which is how much of Australia's reported warming comes from thermometers and how much from the adjustment process. The Bureau's own 2018 research report puts the unadjusted 1910 to 2016 trend at 0.080 °C per decade and the adjusted trend at 0.123, so the adjusted trend is 54 percent higher than the unadjusted one and about a third of the headline figure is adjustment. We checked eight stations we chose ourselves ourselves and found the adjustments raised the trend in 14 of 16 series. That is not proof of anything sinister. It is proof that the uncertainty on the national warming figure is larger than anyone is admitting, and that after sixteen years and three reviews the Bureau still has not published the one number its own 2011 review panel asked for: a statistical uncertainty range on the trend.

---

1. Our verdict first, then the working

We reckon the honest summary is this. The audit found real defects and then oversold them. The Bureau admitted the defects and then undersold them. The Guardian defended the Bureau in a column that reports no check of any number. And all three of them talked past the only question a reader actually cares about: when the Bureau says Australia has warmed by more than 1.5 °C since 1910, how much of that figure would survive if a different, equally defensible set of adjustments had been used?

We do not know. More to the point, the Bureau does not know either, or if it does it has never published it. That is the finding of this paper and everything below is the evidence for it.

A word on where we stand. We run a weather station in Dayboro since 2004 and we publish our own forecasts and our own accuracy scorecards against the Bureau every month. We are not neutral about the Bureau's forecasting, we think ours is better for this valley, we show the numbers, and we sell forecasts on the strength of it. Read what follows with that in mind. On the climate record we have no dog in the fight except one: when a number is used to justify policy that costs real money, the uncertainty on that number must be published alongside it. On this moment it is not.

---

2. This is not the first time. A short history, 2010 to 2026

The Guardian describes sixteen years of "attacks" on the Bureau's temperature record as if the attacks came from nowhere and as if nothing came of them. The record says otherwise. Here is the history as we can source it, and we have left out everything we could not source.

2.1 2010 to 2011: the first audit request

The argument starts in the winter of 2010 on Joanne Nova's site, with a series of guest analyses by Ken Stewart, who sat down with the Bureau's then "High Quality" (HQ) network and compared it, station by station, against the raw data the Bureau also published. His first post, "Australian warming trend adjusted UP by 40%" (29 July 2010), covered the 100 rural HQ sites: the raw records gave a warming of about 0.6 °C per century, the adjusted HQ records about 0.85, and the median adjustment at a site was 0.275 °C. The Bureau's David Jones is quoted in that post, from an email, saying the adjustments have "a near zero impact on the all Australian temperature because these tend to be equally positive and negative across the network". Stewart's figures said otherwise. A second post (14 September 2010) did the same for 34 urban sites and found the raw trend of about 0.4 °C per century became about 0.78 after adjustment, a 70 percent rise when averaged town by town and 95 percent when the series were averaged first. We have not reproduced Stewart's HQ numbers, because the HQ network was withdrawn in 2012, but we note that the direction and size of what he found is the same as what we find in version 2.7 fifteen years later (Section 7.1).

On 24 December 2010 a letter went to the Auditor General, Ian McPhee, asking for a formal audit of the Bureau's temperature record. It was signed by Senator Cory Bernardi, Joanne Nova, Andrew Barnham, Anthony Cox, James Doogue, Chris Gillham, Ken Stewart and David Stockwell, and published on Nova's site on 16 February 2011 together with the Audit Office's reply, which said the request "will be considered as a part of the process of determining our future work program, and within the context of current resourcing and work priorities". No Auditor General review followed. What did follow, the same year, was the Bureau's own Independent Peer Review of the new ACORN-SAT dataset. We read the timing as cause and effect. We cannot prove it, and the Bureau has never said so, so take that as our inference and not as a fact.

The one piece of that early work that made it into the peer reviewed literature is Stockwell and Stewart (2012), published in Energy and Environment. They compared the Bureau's HQ network against a minimally adjusted version of the same stations and found the minimally adjusted trend was 31 percent lower. We note two caveats the Bureau's defenders are entitled to: the journal is not a mainstream climate journal, and the HQ network was replaced by ACORN-SAT the same year so the paper describes a dataset that no longer exists. Both caveats are fair. Neither one answers the finding.

2.2 2011 to 2012: the Independent Peer Review

The Bureau convened an Independent Peer Review panel (chaired by Ken Matthews, with Thomas Peterson, Xiaolan Wang and David Wratt as members) which reported in 2011 with 31 recommendations. The Bureau's response, dated 15 February 2012, agreed to 28 in full and 3 in part. Four of those recommendations matter for what follows and we quote their substance:

• C1: publish the list of adjustments made to each station.
• C2: publish the computer code used to produce the dataset.
• C3: calculate the raw and homogenised national trends by the same method so they can be compared.
• D4: publish a statistical uncertainty range on the national temperature trend.

Hold those four in mind. We will come back to how many of them were done.

2.3 2012: the first ACORN-SAT release and the minimum above maximum problem

ACORN-SAT version 1 was released in March 2012. Within months, external checking found hundreds of days where the homogenised minimum temperature was higher than the homogenised maximum for the same day, a physical impossibility. The check was done by Ed Thurstan, who coded the quality tests the Bureau itself describes in section 6.1 of its own technical report on the dataset (Trewin, 2012) and ran them against the March 2012 release. His write up on Nova's site (16 July 2012) carries 954 in the headline and "about 1,000" in the body, so even the man who counted them gave two figures. Earlier the same year (15 March 2012) a separate check of 8.5 million daily observations at 237 sites had found most of them were not rounded the way the Bureau's own guidance said they should be. We did not try to reproduce either number for version 1 because version 1 is no longer distributed. We did check version 2.7 and found zero such days in 3,970,806 paired observations. So that one got fixed. It took a few years, but it got fixed, and we say so because this paper is not a hatchet job.

2.4 2014 to 2015: Rutherglen, Bourke, Deniliquin and the Technical Advisory Forum

In July 2014 Ken Stewart went through all 104 ACORN-SAT stations and listed six where the adjustments had changed the century trend by more than 2 °C: Brisbane Airport, Amberley, Dubbo, Rutherglen, Rabbit Flat and Carnarvon. In August and September The Australian's Graham Lloyd turned that into a series of front page stories, with Jennifer Marohasy supplying most of the station work. Amberley's minimum trend had gone from about minus 1 °C per century raw to plus 2.5 adjusted, with no recorded site move since 1941; the Bureau said the adjustment came from "neighbour comparisons". Rutherglen, an agricultural research station in Victoria, went from minus 0.35 to plus 1.73. Bourke's minimum trend went from 0.53 to 1.64 °C of warming (Marohasy's figures, which the Bureau did not dispute) and its 51.7 °C reading of 3 January 1909 was removed from the record (the Bureau's reasoning, from a 1997 paper by Blair Trewin, is that a Sunday reading in 1909 was probably a 48 hour maximum, which is fair enough as reasoning goes). At Deniliquin, David Stockwell found a minimum series that cooled 0.7 °C per century raw and warmed 1.0 adjusted, and the Bureau confirmed to him that statistical tests, not documents, were the "primary" justification. The Bureau's public position throughout was that its method was "world's best practice".

Rutherglen became the test case because the people who worked there said the station had never moved. After pressure from The Australian the Bureau released the list of 17 neighbouring stations it had used to detect the breaks. One of them, Hillston, is about 300 km away in a straight line. We looked up what the Bureau's current adjustment list says about Rutherglen, because the list is public now (recommendation C1, see Section 2.2) and in 2014 it was not. Version 2.7 carries eight adjustments at Rutherglen. The two biggest in the minimum column, minus 0.66 °C from 1974 and minus 0.62 °C from 1966, are both labelled "Move" with the comment "Exact details of move unconfirmed". The other two minimum adjustments (2007 and 2016) and all four maximum adjustments are labelled "Statistical". So the Bureau's own file says, in its own words, that the two adjustments that do most of the work rest on a move it cannot confirm.

We did not take either side's trend number. We calculated it ourselves from the version 2.7 files, 1913 to 2025, and the result is in Table 4 in Section 7.1: raw Rutherglen minimum trend minus 0.011 °C per decade, ACORN-SAT 0.196 °C per decade. The adjustment turned a flat series into a warming one. Whether that is right depends entirely on whether the unconfirmed moves happened. Twelve years later we are still waiting on the evidence that would settle it.

The 2014 controversy led to the Technical Advisory Forum being set up in December 2014, a body the 2011 review had already recommended (the Bureau's own 2014 media release says so), with Ron Sandland as its first chair. Its first report (June 2015) found that 18 of the 31 Independent Peer Review recommendations had been fully implemented. It also said something we think deserves more attention than it got. On pages 9 and 10 the Forum records that "no groups other than the Bureau are known to have attempted to produce or analyse an homogenized daily data set for Australia", that "several choices within the adjustment process remain a matter of expert judgment", and then recommends the Bureau limit the help it gives to outside groups trying to replicate the work, or charge for it at cost. Read those three sentences together. Nobody has replicated it, parts of it are judgment, and the Forum's answer was to make replication harder. The Forum's 2017 report recorded that the computer code (recommendation C2) was available "on request" rather than published, that the uncertainty estimate (D4) was still outstanding, and recommended the Bureau establish an "independent assurance mechanism". We have not been able to find that the mechanism was ever established. If it was, we would like to be corrected and we will update this paper.

Two more pieces from 2015 belong here. Chris Gillham compared the Bureau's 1933 and 1953 Year Book temperature tables, the Bureau's own published averages from before homogenisation, against ACORN-SAT and found the adjustments had cooled the period before 1940 by about 0.4 °C; his first headline said this removed half the warming, which he corrected a week later to 30 to 40 percent, and we mention the correction because it is what honest work looks like. A second Gillham comparison across 84 Year Book stations put the raw rise from 1911 to 1940 to 2000 to 2014 at about 0.3 °C against ACORN-SAT's 0.9 at 112 stations. Those are his figures, not ours, and we have not reproduced them. We include them because they are the earliest attempt we know of to do what the 2011 review's recommendation C3 asked the Bureau to do, which is to put the raw and the adjusted national figure side by side by the same method.

2.5 2017: Goulburn, Thredbo and the minus 10 degree floor

In July 2017 observers noticed that the automatic weather station at Goulburn, and later Thredbo, were failing to record very cold readings. On 2 July Goulburn fell to minus 10.4 °C and the value vanished from the Bureau's climate pages. The Bureau's first public explanation was that the reading was being held for quality control. Within the month it had found the hardware cause and commissioned a review. The Bureau's own account, the Review of the Bureau of Meteorology's Automatic Weather Stations (2017), is blunter than most people remember. The MSI1 data card fitted at Goulburn in 2003 and at Thredbo in 2007 could not operate below minus 10.4 °C. It was fitted by field officers who did not know the two card types differed, which the review traces to process and documentation failures dating back to the mid 1990s. The floor bit on one day at Goulburn and on six days at Thredbo since 2007. Cards were replaced at six sites where minus 8 °C had ever been recorded: Goulburn, Thredbo, Tuggeranong, Mount Baw Baw, Butlers Gorge and Fingal. Neither station is in ACORN-SAT. A Senate Estimates answer (SQ22-000099, tabled 22 February 2022, to Senator Rennick) added that 724 automatic stations were reviewed and only those two were affected. The 2017 review panel found the Bureau's data quality control processes "work well, flag errors appropriately" and are in line with practice elsewhere. Two papers on the related question of how fast the Bureau's probes respond, Ayers and Warne (2020) and Ayers (2020), were later published in the Journal of Southern Hemisphere Earth Systems Science. So the narrow claim (a floor on cold readings at two stations) was right and the Bureau confirmed it; the broad claim (a network wide problem) was not upheld. We give both halves because that is what the record says, and because "the Bureau has rejected the claims" turned out, that time, to be half the story.

The same year it emerged the Bureau's automatic stations record a one second temperature sample rather than the one minute average recommended by the World Meteorological Organization, which makes a new "record high" easier to set on a gusty day than it was with a mercury thermometer. The Bureau's position is that its instruments are designed to have a thermal lag comparable to mercury. We have not tested that and cannot verify it either way, so we leave it as an open question rather than a finding.

2.6 2018 to 2022: version 2 and the trend goes up again

ACORN-SAT version 2 was released in late 2018, documented in Bureau Research Report 032 (Trewin, 2018) and later in Trewin et al. (2020) in Geoscience Data Journal. Version 2 raised the national 1910 to 2016 trend from 0.100 to 0.123 °C per decade (Table 9 of the research report, reproduced in our Figure 1). A figure of "23 percent more warming" circulated at the time; 0.123 divided by 0.100 is 1.23, so that number at least is arithmetic, not opinion. The series has been revised and extended several times since. The October 2026 release is version 2.7.

2.7 2026: the Berger and Sherrington audit

Which brings us to now. The audit document (we read all 40 pages) compares the Bureau's Climate Data Online "raw" archive against ACORN-SAT version 2.6. Its author describes the method only as "statistical software and computer coding" plus a Monte Carlo appendix. There is no methods section. That is a real weakness and Readfearn is right to say so. It does not make the findings wrong; it makes them unverified until someone verifies them. So we did.

The audit reached the public through The Australian, two articles and an editorial, and through a long post on Joanne Nova's site on 17 October 2026 which is where most readers will have met the numbers: around 150 stations affected, 879 "value traps" (runs of identical readings), over 500 "structured repeats" (copied blocks), a 647 day gap at Sydney, and the Sydney maximum adjusted in one direction for 110 years. That post also published the Bureau's reply to The Australian, from its Chief Executive Stuart Minchin to Andrew Wilkie MP, which "rejects allegations of fraudulent or deliberate manipulation" as having "no basis in fact" and says the "duplicated data affects fewer than 200 daily values from a dataset with more than seven million values". Berger's answer, in the same post, was that the Bureau "acknowledged a few trees but denied the existence of the forest". Both of those sentences can be checked against the file. We do that in Section 3.1.

The loudest version of the critics' case came on 18 October, when Peter Ridd, interviewed on the Outsiders programme, said the audit showed a "general complete lack of quality assurance in the Bureau of Meteorology" and that the checking should be done by a hundred scientists rather than "these two gentlemen who've actually done it for free" (Ridd, 2026). He also named Jennifer Marohasy, Bill Johnston and John McLean as having raised the same kind of problem years earlier, which matches the history above. That is the strongest reading on offer. Ours, in Section 3, is narrower: some of the audit's findings hold up against the file and some do not, and the interesting question is which.

---

3. What the audit claims, and what we found when we checked

We checked every claim that coudl be checked against the raw files in the Bureau's own version 2.7 bundle. The results are in the sections below. The short version: the data findings hold up, the interpretation does not always.

3.1 The duplicated months at Sydney Observatory Hill

The audit says six pairs of months in the Sydney Observatory Hill (station 066062) minimum temperature record are exact duplicates of each other. We opened the file and compared them day by day.

Pair of months	Days compared	Identical minima	Identical maxima
July 1864 and July 1865	31	31	0 to 4
August 1867 and August 1907	31	31	0 to 4
May 1883 and May 1884	31	31	0 to 4
November 1903 and November 1905	30	30	0 to 4
July 1914 and July 1915	31	31	0 to 4
June 1943 and June 1944	30	30	0 to 4
December 1935 and December 1936	31	29 (days 15 and 16 differ)	not checked
April 1869 and April 1870 (our scan)	30	30	0
August 1868 and August 1869 (our scan)	31	2	30 (day 2 differs)
June 1865 and June 1905 (our scan)	30	0	29 (one day missing in 1865)

Table 1. Sydney Observatory Hill raw file hqnew066062 from the ACORN-SAT v2.7 bundle, file dated 29/04/2021. Comparison by Dayboro Weather Research. "0 to 4" means the same pairs of months in the maximum temperature column match on at most four days, which is what you expect from chance. The first seven rows are the pairs the audit names. The last three are pairs the audit does not mention which turned up when we scanned every same month pair in the file for 27 or more matching days.

Six exact duplicates of a full month and one spliced month at Sydney in the minimum column, with the maximum column for the same months behaving normally. Then we went looking for what the audit missed, and found three more: a seventh exact minimum pair (April 1869 and 1870) and two near complete pairs in the maximum column, August 1868 against August 1869 and June 1865 against June 1905, each with the other column behaving normally. So the copying is not confined to the minima, it just happens to be rarer in the maxima, and a 40 year gap turns up in both columns. That is not a weather coincidence. Two Julys in Sydney do not produce 31 identical daily minima to a tenth of a degree. Figure 4 shows one pair.

Figure 4. Sydney Observatory Hill daily minima, July 1914 (solid) and July 1915 (dashed). All 31 values identical. Source: BOM ACORN-SAT v2.7 raw bundle.

The December 1935 and 1936 pair is more interesting than a plain duplicate. Days 1 to 14 match, days 17 to 31 match, days 15 and 16 do not. The audit calls this a "seam", meaning one month was patched into the other with a two day splice in the middle. We confirm the pattern. We cannot tell you how it happened. Neither, on the evidence of its statement, can the Bureau.

Charleville (044022), December 1942 against December 1943: by our count 29 of 31 days identical in the minimum column (days 15 and 17 differ) and 29 of 31 in the maximum column (days 1 and 14 differ). The audit describes this pair differently, as two copied blocks with a seam in the minima and 13 matching maxima; we report what the file shows.

Now the two Bureau counts. The Guardian quotes a Bureau spokesperson saying ten months at Observatory Hill showed duplicate minima, three of them after 1910. Our eight Sydney minimum pairs are sixteen months, three pairs of them after 1910. If the Bureau counts the copied month only (one per pair) and leaves out the 1864/65 pair as falling before the official record, or leaves out the 1935/36 seam because it is not an exact copy, you can get to ten several ways, and the "three after 1910" matches our three post 1910 pairs exactly. So the ten is plausible, but the Bureau has not published the list and it should. The second count is the Chief Executive's "fewer than 200 daily values". Count one copied value per duplicated day in the six exact minimum pairs the audit named and you get 184. That is under 200, so the sentence is true on that definition. Add the April 1869/70 pair (30), the December 1935/36 seam (29), the two maximum pairs (30 and 29) and Charleville's two columns (29 and 29) and the same definition gives 360. On the Bureau's own framing that is still a rounding error in seven million values. On ours the question was never the share, it was that none of it was caught for sixty years by a quality system the Bureau describes as world class, and that the file is still labelled raw.

3.2 These errors are still in the "raw" data the Bureau published this month

This is the finding we did not expect and it is, we think, the most important one in Section 3. The Bureau's 22 October statement says the duplicated values were "corrected in this year's update, completed on 19 October 2026". We downloaded that update. The homogenised ACORN-SAT files have the affected months blanked on one side of each pair: Sydney July 1915, June 1944 and December 1935 are empty while July 1914, June 1943 and December 1936 are present, and Charleville December 1943 is empty while December 1942 is present. So the four post 1910 months the Bureau refers to are, on our reading, these four, and "corrected" means "deleted from the homogenised series".

But the raw station files shipped inside the same bundle, the files a researcher would use to check the Bureau's work, still contain every one of the duplicates. The Sydney raw file is dated 29 April 2021, the Charleville file 7 December 2009. Nobody touched them. The Bureau's statement says "original observations are retained separately, quality issues are recorded, and ACORN-SAT is produced as an additional analysed dataset". Taken at its word that is consistent: the raw file is kept as it was, the fix lives in the homogenised file. Our point is narrower. If the raw file contains a transcription error from the 1960s, as the Bureau suggests, then the raw file is not the original observation, it is a 1960s copy of it with known errors, and on this moment nothing on the download page tells a researcher that. The "quality issues are recorded" somewhere; we could not find where.

We also note the Bureau's method of correction. The audit found the duplicate but could not say which month was the real one and which the copy. The Bureau blanked the later month in each pair. It did not explain why the later month, and we could find no document that says the field books were consulted. Perhaps they were. It would take one sentence to say so.

3.3 Palmerville: the audit is wrong about the odds and right about the data

The audit's headline example of "human intervention" is Palmerville (028004) on Cape York, where the maximum temperature reads 31.0 °C for 17 consecutive days from 21 April to 7 May 2004. The audit puts the chance of this at ten to the minus 34 and calls it "less likely than winning the lottery every week for a year". We confirm the 17 days (Figure 5). We do not accept the odds.

Figure 5. Palmerville daily maximum, March to June 2004. The shaded block is 17 consecutive days at 31.0 °C. Every value in this window is a whole degree. Source: BOM ACORN-SAT v2.7 raw bundle.

Look at the whole chart, not just the shaded part. Every single value from March to June 2004 is a whole number. The Bureau's response says the site recorded to the nearest degree at the time, and the file bears that out: all 122 valid maximum readings from March to June 2004 are whole degrees, 364 of the 365 maxima in 2002 are whole degrees, 311 in 2003 and 289 in 2004, and the longest unbroken run of whole degree maxima is 312 days from 21 December 2001. From 2005 the file goes back to tenths. A thermometer read to the nearest degree, in the dry season on Cape York where the daily maximum sits in a narrow band for weeks, will produce runs of identical readings. The audit does not say what precision or what day to day spread its probability assumed, but a figure like ten to the minus 34 is only reachable if you treat each day as an independent draw at tenth of a degree resolution. Neither holds for that station in that year. We have not computed the correct odds ourselves, so we say only that the Bureau's explanation is plausible and the audit's number is not. The audit should have caught this, because the integers are right there in the file. This is the kind of mistake you make when you do not have a methods section, and it handed the Bureau an easy win.

What the Bureau's reply did not mention is the gap. Palmerville has no data at all from 17 October 1999 to 26 July 2001, which we count as 648 days (the audit says 647, the one day difference is a counting convention and does not matter). A station with a 648 day hole and several years of whole degree readings is in ACORN-SAT as one of the 112 "carefully chosen locations that maximise both length of record and network coverage across the continent". Coverage, we accept, is why it is there; Cape York has few long records. We would like to know what the quality threshold for inclusion is, because this station does not look like it meets one.

Cape Otway (090015) shows the same integer pattern from 1995 to 2004. The audit reports it. The Bureau did not address it. We confirm it from the file.

3.4 Sydney: every year adjusted upward

The audit's most striking claim, and the one neither the Bureau nor the Guardian touched, is that the ACORN-SAT maximum temperature series for Sydney sits above the raw series in every single year. ACORN-SAT lists Sydney under the current site number 066214, built from the Observatory Hill record 066062 up to the site change on 17 October 2017. We calculated the annual mean of ACORN-SAT minus the annual mean of raw 066062 for every year from 1910 to 2017, 107 complete years. The difference is positive in 107 of 107. The audit says 110 of 110 over 1910 to 2020; we stopped at the site change because after it the two files describe different instruments, and the result is the same either way. Figure 3 shows the decade means.

Figure 3. Sydney Observatory Hill, ACORN-SAT v2.7 minus raw, decade mean of the annual difference, 1910 to 17/10/2017. Calculation by Dayboro Weather Research.

Now, we have to be careful here, and this is where the audit was not. An adjusted series sitting above the raw series in every year is not, by itself, evidence of a warming bias in the trend. If the adjustment was the same size every year it would shift the line up and leave the slope alone. What matters for the trend is whether the adjustment is bigger in the early years than the late years. It is. The 1910s are adjusted up by 1.45 °C and the 2010s by 0.37 °C. The correct reading is therefore not "the adjustments make Sydney warmer", it is "the adjustments make the Sydney past warmer relative to the present, which reduces the Sydney warming trend". For maximum temperature at this one station the adjustment works against the warming story, not for it.

We say this because it is true and because it shows what an honest analysis looks like. The audit presents 110 of 110 as a red flag. On its own it is not. It is proof that the Bureau believes the old Observatory Hill site read cool for a century, which is an arguable position. The adjustment list for Sydney in version 2.7 has eleven entries; the largest is a 1.10 °C step in January 1917 attributed to a site move, and a minus 0.44 °C step in 1938 attributed to a building. Those are the adjustments the Bureau should be asked to justify with overlap measurements. "Every year is positive" is the wrong question.

3.5 The machine learning claim and the 50 largest events

The audit reports that a machine learning model trained on the adjustments could predict them with an R squared around 0.70, and offers this as evidence that the adjustments follow a rule rather than the metadata. We could not reproduce this because the audit does not say what model, what features or what split was used. We make no finding on it. Similarly, the audit's list of the 50 largest "value trap" events across 17 stations (Warruwi 9, Montague Island 4, Palmerville 3) is a list we could not independently regenerate without the audit's definition of an event. These are the parts of the audit that most needed a methods section and did not get one.

---

4. The Bureau's response, sentence by sentence

The Bureau's statement of 22 October 2026 is 500 words. We take the substantive sentences in turn.

"These claims are false and unsubstantiated." The claim of deliberate manipulation is unsubstantiated; we agree, the audit does not demonstrate intent and could not. The claim that the record contains duplicated months is substantiated, by the Bureau's own next paragraph. Leading with "false" and following with "we confirmed four months of it" is the kind of sentence a press office writes, not a scientist.

"Fewer than 200 daily values from a dataset containing more than 7 million values." True and beside the point. Nobody argued that four duplicated months change the national trend. The argument is that a dataset with no methods published, that passed three reviews, still had full month duplications in it that two private citizens found with off the shelf statistical software. The number that matters is not 200 over 7 million. It is how many other classes of error of this kind exist that nobody has looked for yet. The Bureau says it looked and found "no further duplicate months in the entire bureau network". We accept that for duplicate months. It says nothing about the seam at December 1935, which is not a duplicate month and which the Bureau did not mention.

"They are all from handwritten records of pre-1946 data that have been manually digitised." Agreed, and this is the strongest point in the Bureau's favour. Transcription errors in hand punched records are mundane. We would only ask why, if the field books still exist, the Bureau did not go back to them before blanking the later month in each pair.

"Lower variability of temperature in places such as the tropics which make repeated patterns in raw data more likely." Correct, as we show in Section 3.3. The audit was wrong on Palmerville's odds.

"Independently reviewed in 2011, 2015 and 2017, with those reviews expressing overall confidence." They did. They also made 31 recommendations, of which the 2015 review found 18 fully implemented, and the 2017 review found the uncertainty estimate still not done and the code still not published. "Overall confidence" and "among the best in the world" are the sentences from those reviews the Bureau quotes. "Please publish an uncertainty range" is the sentence it does not.

"The bureau does not overwrite or delete original temperature observations." See Section 3.2. The raw file still has the duplicates; the homogenised file has them blanked. That is consistent with the sentence. It also means the file labelled raw carries known errors and the download page does not say so.

What the statement does not address at all: the one way adjustment at Sydney, the machine learning finding, the Cape Otway integers, the 648 day Palmerville gap, and the December 1935 seam. Five of the audit's substantive findings, zero words.

---

5. The Guardian piece: where is the evidence, asks a column that reports none of its own

Graham Readfearn's column of 25 October is headlined "Where is the evidence for those big claims?". It is a fair question. So we counted what evidence the column itself contains. We can only judge what was printed, not what its author read before writing it.

• Number of data points from the audit, the raw files or ACORN-SAT that the column reports checking: zero.
• Number of scientists quoted on the substance: zero. The only person quoted is Andrew Wilkie MP, who forwarded the audit to the Bureau and says he "felt uncomfortable" about how The Australian framed his role. Wilkie's discomfort is a fact about Wilkie. It is not evidence about the data.
• Number of the audit's findings engaged with on the merits: two, the duplicated months (conceded, then dismissed as before 1910) and Palmerville (correctly rebutted with the Bureau's integer explanation).
• Number of the audit's findings not mentioned: the Sydney one way adjustment, the December 1935 seam, the Cape Otway integers, the 648 day gap, the machine learning result. Five. (Charleville gets one sentence, as a count, with no detail.)

Beyond those two points the column's method, as far as we can see, is to describe the authors. One is a corporate entertainer who "mesmerized corporate audiences". The other signed a declaration in 2022 about CO2. Both things may be true, and neither makes 31 of 31 identical July minima less identical. A finding is not refuted by the CV of the person who made it. If it were, the Bureau's finding that there are 200 bad values would be refuted by the fact that the Bureau has an institutional interest in there being only 200.

Three specific claims in the column deserve a closer look.

"Data before 1910 is not used in ACORN-SAT and so, in the context of what the bureau's data says about Australia's changing climate, it is irrelevant." This is the argument several commenters have also made: 1867 and 1907 predate the record, so who cares. On the national trend, fair enough. Two answers all the same. First, four of the duplicated pairs are after 1910 (July 1914 and 1915, June 1943 and 1944, December 1935 and 1936 at Sydney, December 1942 and 1943 at Charleville), as the Bureau itself confirmed and as the column itself reports two paragraphs later, so the pre 1910 defence covers less than half the finding. Second, the pre 1910 Sydney record is used, constantly, whenever a Sydney temperature is described as the hottest or coldest since records began. If the record is good enough for a headline it is good enough to be audited.

"The ACORN-SAT has already been the subject of two independent reviews." The Bureau's own statement says three. Minor, but it suggests the column was not checked against the statement it was defending.

"Attacks on the bureau's temperature records have been orchestrated by climate science deniers for more than 15 years. Previous episodes have been described by scientists as 'harassment'." No scientist is named, no episode is named, and no source is given. Section 2 of this paper is the fifteen years. The 2017 Goulburn episode, where the "attack" turned out to be a hardware fault the Bureau confirmed in a Senate answer and two published papers, is the clearest test of whether these episodes find things. It found a card that could not record below minus 10.4 °C fitted to automatic stations whose operators had not been told. If that is harassment, the Bureau should wish for more of it.

Well, we would put one question to the column. Does it say whether anyone opened the raw file? It does not. The file is a public download. It takes ten minutes to confirm that July 1914 and July 1915 are identical and about one hour to confirm they are still identical in the October 2026 bundle that the Bureau says fixed them. The column runs to well over a thousand words. One of them could have been a number.

---

6. The other arguments doing the rounds

Several further defences of the Bureau were put to us in discussion over the past week. We take them seriously and answer each.

6.1 "UNSW scientists already showed the Bureau is not fiddling the data"

The article in question, by Andy Pitman and Lisa Alexander, was published on 1 September 2014 during that year's argument over homogenisation, and the dataset it describes is ACORN-SAT version 1. Its one data point is worth repeating: the authors say their own work found larger trends in temperature extremes in the unadjusted data than in the adjusted. That is a real finding and it cuts against the simple "fudge" story for extremes. It does not touch the mean trend in Table 9, which goes the other way. Version 1 was replaced by version 2 in 2018 and the trend rose by 23 percent in the process. An article that defends a dataset which its own authors have since replaced with one that warms faster cannot be a defence of the current dataset. It is a historical document. We do not say the authors were wrong in 2014. We say 2014 was twelve years and two major revisions ago.

6.2 "Berkeley Earth, NASA and NOAA independently confirm Australia's warming"

Three groups do report a similar Australian trend. Whether that is independent confirmation depends on where their Australian data comes from. Berkeley Earth's published source list for its land record is built mostly from the Global Historical Climatology Network and related archives, and for Australia those archives are fed by the Bureau's own observations, because nobody else runs the thermometers. We could not find any Australian temperature series built from instruments the Bureau does not operate. If one exists, eventually we have missed it, and we would like to see it. Until then "three datasets agree" means "three groups processed the same thermometers and got similar answers", which is reassuring abotu the processing and says nothing about the thermometers. We mark this as a point we have not fully verified: we traced Berkeley Earth's station sources only as far as the Global Historical Climatology Network, we did not research the NASA and NOAA products at all, and the full provenance chain is something we intend to trace in a follow up.

6.3 "200 errors in 7 million is 0.003 percent, so it does not matter"

We expect this to be the first reply from the Bureau's defenders, so we deal with it in full rather than in passing. It is true as arithmetic. Here are the numbers, worked three ways.

Count	Divided by	Rate
Fewer than 200 duplicated values (the Bureau's count)	About seven million daily values in the network	0.003 percent
360 values (every pair we found, Section 3.1)	About seven million	0.005 percent
16 duplicated months at Sydney Observatory Hill	About 1,900 months of Sydney record	0.8 percent

Table 2. The same errors stated as a rate three ways. The denominator is a choice, and each choice is an argument. Source: Bureau statement 22 October 2026 for the first row, our scan of the v2.7 raw files for the other two. Calculation: Dayboro Weather Research.

So yes, the duplicated months were never going to move the national trend, and anyone who says they prove the warming is made up has overreached. We say that plainly in Section 3 and we say it again here. But the percentage answers a question nobody asked. The audit did not claim the errors change the trend. It claimed the Bureau never checked. Five reasons the rate is the wrong measure.

The size of the error is not the finding. Not finding it is. The Bureau has told two reviews and the Senate that its quality control processes "work well, flag errors appropriately" (Bureau of Meteorology, 2017). A check for a copied month (does this month equal any other month at the same station?) is about the simplest test there is, and the copies sat in the published file for decades until two retired people ran it in their own time. The Bureau has not said when, if ever, that check was last run before October 2026, and the reviews that expressed "overall confidence" did not report running it. The question is not what these 200 values did. It is what else the same net let through, and nobody can answer that, the Bureau included.

The errors sit in the file labelled raw, and raw is the control. Every adjustment in ACORN-SAT is justified by comparing a station against its neighbours' unadjusted records. An error in a reference station does not stay in its own cell. It becomes a reason to adjust somebody else, and the size of that second effect is not known. Small input, uncertain output, and the uncertainty is unquantified. That is the 2011 review's open item, fifteen years on (Section 7).

Any rate is a floor. The audit found the class of error a pattern search can find: identical runs and copied blocks. A 23.4 keyed as 32.4 leaves no pattern. A month shifted by one day leaves almost none. So a rate of detected errors is a lower bound on the real rate, and the gap between the two is the thing we would most like to know.

The Bureau's own count moved. First "fewer than 200 values", then "ten months", and our scan found more pairs than the audit named (Section 3.1). If the rate is too small to matter it should at least be stable. A number that changes each time somebody looks is not a number anybody has measured.

Small is a comparison, so compare it with the right thing. The whole gap between the adjusted and unadjusted national trend is 0.043 °C per decade, about a third of the headline warming (Table 3). Everything in this argument lives in small numbers. If a few hundred values are too small to worry about, so is the adjustment that lifts Sydney's 1910s maximum by more than a degree (Figure 3), and the Bureau defends that one to the decimal. You cannot run "small does not matter" in one direction only.

The percentage tells you the errors found so far are small. It tells you nothing about the errors not found yet, and the way these ones were found suggests nobody was looking.

6.4 "Correct every error the audit found, recalculate the trend, and you will see it makes no difference"

We agree that it would make no difference, and the Bureau has effectively done it by blanking four months. But this misreads what the audit is for. The 200 values are not the claim. The claim is that the adjustment process, which moves the national trend by 0.043 °C per decade, is not independently verifiable because the inputs to it (station metadata, overlap measurements at moves, the statistical detection thresholds) are not published in a form that lets anyone recalculate it. Fixing 200 values does not touch that. See Section 7.

6.5 "The adjustments go both ways, so there is no bias"

This one we checked, because it is the Bureau's standing defence and it is partly right. The version 2.7 adjustment list has 1,015 entries: 465 positive, 547 negative, 3 of zero size. By count, more adjustments are negative. Figure 6 shows them by decade.

Figure 6. ACORN-SAT v2.7 adjustments by decade of effect, 1,015 entries. Count by Dayboro Weather Research. The sign convention of the "Size" column is as listed by the Bureau; we have not found a Bureau document that states it explicitly.

But a count is the wrong measure, for the same reason "110 of 110" was the wrong measure at Sydney. An adjustment changes the trend according to when it is applied and in which direction. A negative adjustment applied to the years before a break cools the past and raises the trend. A negative adjustment applied to the years after a break cools the present and lowers it. The list does not tell you which, and the Bureau's own Table 9 tells you the net effect: the adjusted trend is higher than the unadjusted trend for mean, maximum and minimum temperature, over both 1910 to 2016 and 1960 to 2016. The count of signs is a distraction. The net effect is what the data says and it goes one way.

---

7. The question that matters: how much of the warming is thermometer?

This is the part we most want you to read. Everything above is about whether the Bureau makes mistakes. Every organisation makes mistakes. The question is whether the headline number survives them, and the Bureau's own research report is the best source we have.

Trewin (2018), Bureau Research Report 032, Table 9, gives the 1910 to 2016 trend for three versions of the same network:

Series	Mean	Maximum	Minimum
Unadjusted (AWAP gridded, Bureau network)	0.080	0.068	0.091
ACORN-SAT version 1	0.100	0.090	0.109
ACORN-SAT version 2	0.123	0.116	0.130
Version 2 minus unadjusted	0.043	0.048	0.039
Version 2 as a multiple of unadjusted	1.54	1.71	1.43

Table 3. Linear trend 1910 to 2016 in °C per decade. First three rows: Trewin 2018, BRR032, Table 9. Last two rows: our arithmetic from those values.

Figure 1. One national network, three national trends. The unadjusted series is the Bureau's gridded AWAP product, which uses a wider station set than ACORN-SAT. Source: Trewin 2018, Table 9.

Read that table slowly. The thermometers, unadjusted, say 0.080 °C per decade. The Bureau's current adjusted series says 0.123. Over 106 years that is the difference between 0.85 °C of warming and 1.30 °C. The adjusted trend is 54 percent higher than the unadjusted one; put the other way, about a third of the adjusted figure is adjustment. The Bureau has never disputed this; it is the Bureau's own table. One caveat the report itself makes: the unadjusted row is a gridded product on a wider station set, not the same 112 stations run through the same code, so part of that gap may be station choice rather than adjustment. That is exactly what recommendation C3 asked the Bureau to separate out. What the Bureau says is that the adjusted number is the right one because the raw network has known cooling biases (sites moving out of towns, screen changes) that homogenisation correctly removes. That may be so. It is a hypothesis, and the research report itself gives you the reasons to hold it loosely:

• Version 2 has 966 adjustments (463 maximum, 503 minimum) against 660 in version 1. Of these, 513, or 53 percent, are supported by station metadata (Trewin 2018). The other 47 percent were detected statistically, meaning the series was compared against neighbouring stations and a step was found. Table 7 of the report lists 209 maximum and 244 minimum adjustments as "statistical" in cause. Nearly half the adjustments in the national climate record have no documentary evidence that anything happened at the station.
• Only 91 of the 966 adjustments, about 9 percent, were made using overlapping observations from the old and new instrument (Table 2 of the report). The other 91 percent were estimated from neighbours. An overlap is the only direct measurement of what a move or a screen change actually did. For nine in ten adjustments there is none.
• Version 1 had a rounding bias of about plus 0.025 °C per adjustment, which the report says accumulated to roughly 0.08 to 0.09 °C per station over the record. This was not found by a reviewer. It was found when the Bureau rebuilt the dataset six years later. Section 7 of the report describes version 1 as having had "two approximately offsetting biases". The sceptic question is obvious: how many offsetting biases does version 2 have that will be found in version 3?
• The 2011 review's recommendation D4, a statistical uncertainty range on the trend, has still not been met in the form the panel asked for. The report discusses sources of uncertainty. It does not say "0.123 plus or minus X". The 2017 Forum noted it as outstanding. We have searched the Bureau's ACORN-SAT pages and the 2018 report for a published trend with a confidence interval since then and have not found one. If it exists, we will link it.

7.1 We checked eight stations ourselves

A national table can hide a lot, so we picked eight ACORN-SAT stations ourselves. We chose them, not at random, because they have long records and several have featured in past controversies, so this is a check on the famous cases, not a sample of the network. Sydney is not in the table because Section 3.4 already shows its maximum adjustment lowers the trend. For each we computed the annual mean trend from the raw component files and from the homogenised v2.7 file over the same years. Table 4 and Figure 2.

Station	Years	Max raw	Max ACORN	Min raw	Min ACORN
Rutherglen	1913 to 2025	0.095	0.159	-0.011	0.196
Bourke	1911 to 2025	0.057	0.098	0.071	0.146
Darwin	1910 to 2025	0.029	0.192	-0.014	0.094
Amberley	1942 to 2025	0.169	0.158	-0.056	0.217
Cape Otway	1910 to 2025	0.093	0.086	0.088	0.114
Alice Springs	1910 to 2025	0.156	0.230	0.033	0.139
Deniliquin	1910 to 2025	0.087	0.129	-0.069	0.091
Marble Bar	1910 to 2025	0.004	0.095	0.098	0.130

Table 4. Linear trend of annual mean temperature, °C per decade, raw component sites against ACORN-SAT v2.7, years with at least 300 valid days. Calculation by Dayboro Weather Research from Bureau files. Marble Bar and Bourke raw series are composites of the component sites the Bureau lists for each ACORN station.

Figure 2. ACORN-SAT v2.7 trend minus raw trend at eight stations. Positive means homogenisation increased the warming trend.

Fourteen of sixteen series warm faster after adjustment. The two that do not (Amberley and Cape Otway maximum) move down by 0.011 and 0.007 °C per decade. Four raw minimum series (Rutherglen, Darwin, Amberley, Deniliquin) show cooling over a century and the adjusted versions of all four show warming of 0.09 to 0.22 °C per decade. At Amberley, an air force base west of Brisbane, the minimum trend goes from minus 0.056 to plus 0.217, a swing of 0.27 °C per decade, on a site the Bureau itself lists as a single component station for its whole record.

We want to be exact about what this does and does not show. It does not show the adjustments are wrong. Rural minimum temperatures are the variable most affected by site moves out of town, by screen changes and by vegetation, and there are physical reasons an unadjusted rural minimum would read too cool over time. It shows that at these eight stations the adjustments are large, they go one way for the trend in fourteen of sixteen series, and at several stations they are the entire warming signal. Two cautions on our own method: where the Bureau lists several component sites we joined them end to end without any overlap correction, which can itself bias the raw trend, and Darwin's documented move from the post office to the airport is exactly the kind of break homogenisation exists to fix. Any reader who wants to argue Rutherglen has warmed must, in the end, be arguing that the undocumented site move happened and was of exactly the size the neighbours imply. That is a defensible argument. It is not an observation.

---

8. Why a national agency should not be making these mistakes

We have been fairly hard on the audit in this paper. We will now be equally hard on the Bureau, because the standard it is held to is different. The audit was written by two private citizens with statistical software. The Bureau is a Commonwealth agency with a climate data section, a budget, three external reviews, and it quotes those reviews as rating its practices "among the best in the world". Four things follow from that.

First, a month copied into the wrong year is a solved problem. Duplicate detection is a one line check. The Bureau's own statement says it "conducted analysis of all underlying data" and found no further duplicate months in the whole network, which means the check exists and runs in hours. It was run in October 2026 after a newspaper asked. The Bureau has not said whether it was ever run before, and the duplicates sat there through sixty years of digitised records and three reviews. If the check had been run, the result was not acted on. If it had not, nobody thought to. Either way it is not a transcription error. It is a quality control gap, and it was an outsider who closed it.

Second, a dataset labelled "raw" must be raw or must say what it is. We showed in Section 3.2 that the raw files in the October 2026 bundle still carry the duplicates. If the field books are the originals and the digital file is a 1960s copy with known errors, the download page should say "digitised copy, known transcription errors, see list". Then a researcher knows what they have. On this moment they do not.

Third, the uncertainty has been requested since 2011 and has not been delivered. Every other measured quantity in science carries an error bar. The Bureau's "more than 1.5 °C since 1910" does not, in the public material, and the research report that underpins it shows the number moving by 54 percent depending on method. A panel the Bureau itself appointed asked for this fifteen years ago. The Technical Advisory Forum noted it was still missing in 2017. A national agency that cannot, or will not, publish the uncertainty on its headline number after fifteen years of being asked has not earned "overall confidence". It has earned a question.

Fourth, how the first sentence is written costs trust. Goulburn 2017: the first word was quality control, then a hardware floor was confirmed in the Bureau's own review and to the Senate. Duplicated months 2026: the statement opens with "false and unsubstantiated", aimed at the manipulation claim, and the reader has to get to the second paragraph to learn that the duplicates themselves are confirmed. Both things in that statement are true. The order they come in is a choice, and it is the choice a press office makes rather than a data section. Each time the concession is small and the Bureau is right that the trend survives. A reader who has watched a few of these does not need to believe in a conspiracy to start reading past the first sentence.

---

9. What we do not know, and four questions for the Bureau

Here is what this paper cannot tell you. It cannot tell you that the Bureau's adjustments are wrong. It cannot tell you Australia has not warmed; our own Dayboro record, short as it is, shows warming and we have said so on this site. It cannot tell you that the 0.043 °C per decade of extra trend is bias rather than correction. What it can tell you is that nobody has published the information that would let you decide, and the Bureau is the only party that has it. So we put four questions, and we will publish any answer the Bureau gives, in full.

1. Of the 0.043 °C per decade difference between the unadjusted and the version 2 trend, how much comes from the homogenisation adjustments themselves and how much from the choice of which stations and which periods to include? Table 9 of the research report compares ACORN-SAT against AWAP, which is a gridded product on a different station set. Recommendation C3 of the 2011 review asked for a like for like comparison. The Bureau has said it addressed C3 in its 2012 technical report (Fawcett et al., 2012). We read that report. It compares ACORN-SAT against two whole network gridded analyses, AWAP and an unadjusted set the authors call WNDC, and finds the homogenised maximum series sits about 0.2 °C below the unhomogenised ones at the start of the record, a gap that closes by about 1960. That is a comparison of networks, not of the same stations, and the report says the uncertainty from undetected early discontinuities "will be the subject of further study". The 2017 Forum still listed C3 as outstanding. One table, unadjusted against adjusted on the same stations and the same years, would settle it.
2. For the 50 largest adjustments in version 2.7, what measurement validates each one? The report says 91 adjustments used overlapping observations. Which ones? For the remaining 875, what is the neighbour set and what is the detection threshold? A table with one row per adjustment answers this, and the adjustment list already exists, it only needs the extra columns.
3. What is the statistical uncertainty range on the 1910 to 2025 national trend, including the uncertainty from adjustment choices? Recommendation D4, 2011. Noted as outstanding by the 2017 Forum, and we have not found it published since.
4. What does the national trend come out at under other defensible homogenisation methods, and has a blind test ever been run? Venema et al. (2012) ran a blind benchmark on European methods and found that homogenisation generally improves trends but that the methods disagree with each other. Lund et al. (2023) and Killick et al. (2022) have since set out how to test changepoint detection against known answers, and in the United States Hausfather et al. (2016) checked the adjusted network against a separate reference network of pristine sites, which is the kind of outside check Australia does not have. The Bureau's method, like the American one described by Menne and Williams (2009) and benchmarked by Williams, Menne and Thorne (2012), leans on comparison with neighbouring stations, so benchmarks of that kind exist for it. Has ACORN-SAT's method ever been put through a blind test with synthetic breaks, and if so what did it score?

None of these questions assumes bad faith. All of them have answers that fit on one page. The fact that sixteen years of argument have not produced that page is the finding.

---

10. Conclusion

The October 2026 audit found real errors in the Bureau's raw temperature archive: six duplicated months at Sydney, one at Charleville, a spliced month, and long stretches of integer only readings at two reference stations. We verified all of these against the Bureau's own current files. The audit also made claims it could not support, about Palmerville's odds and about human intent, and the lack of a methods section means its larger statistical claims remain unverified. The Bureau's response conceded the duplicates, correctly explained Palmerville, and left the Sydney adjustments, the seam, the gap and the integers unmentioned, while shipping a "raw" bundle that still contains every error it says it fixed. The Guardian's defence quoted no scientist and reported checking no number.

Behind all of it sits the Bureau's own Table 9: 0.080 °C per decade from the thermometers, 0.123 after adjustment. We do not know how much of that gap is correction and how much is method, and after three reviews and fifteen years of asking, neither does the public. The Bureau says the lower number carries biases of its own and the higher one is the better estimate. It may well be right. But until it publishes an uncertainty range on its headline trend and the validation behind its largest adjustments, "Australia has warmed by more than 1.5 °C since 1910" is a number without an error bar, and the Bureau's own table shows the method can move it by a third. We have not been told how much of that third is thermometer.

We will keep checking. The files are public, and that, at least, the Bureau deserves credit for.

---

Related Reading on Dayboro.au

---

References

Data files used for every calculation in this paper (annual means, station trends, duplicate month comparisons and adjustment counts) were produced from the Bureau's ACORN-SAT v2.7 public release and are available on request.