Summary of Key Findings: BOM ACORN-SAT Data Integrity Audit Prepared by: Thomas Berger and Geoff Sherrington Date: October 2026 Subject: Systemic data integrity issues in the Bureau of Meteorology's ACORN-SAT temperature dataset --- 1. Whole-Month Replications at Sydney (Station 066062) The audit identified multiple instances of entire months of daily minimum temperatures being replicated exactly in a different year, all flagged Quality = Y by the Bureau's own quality control. - June 1943 and June 1944: All 30 days identical. This falls inside the pre-1946 window the BOM itself cites as the source of duplicates, so the "pre-1946" explanation does not dispose of it — it includes it. - November 1903 and November 1905: All 30 days identical. (Pre-1910; included for completeness.) - January/February 1910: A seven-day block replicated across the January/February boundary, in the first year of the Bureau's standardised record. Why it matters: Two complete months of daily observations cannot both be original. One is a copy of the other. The QC flag cannot be correct for both. The BOM's statement says it has "detected no further duplicate months in the entire bureau network other than those quoted in the report and then confirmed by us." That sentence is contradicted by the BOM's own published data. --- 2. The 1935/1936 Seam A different structure appears in December 1935 and December 1936: - 14 days identical (Dec 1–14). - Two days present only in 1936 (Dec 15–16). - A shared value of 20 on Dec 17, with the 1935 record carrying no days-of-accumulation flag. - 14 more days identical (Dec 18–31). Why it matters: This is the signature of a block splice — a join point with a two-day insert and a malformed record at the stitch. It is not an isolated transcription error; it shows the mechanism by which blocks were joined. --- 3. Scale of Duplication The audit identified over 1,000 duplicated blocks, with block sizes down to five days. 42% of duplication events occurred post-1990 — in the modern digital era, not the pre-1946 handwritten era the BOM cites. - The audit did not analyse the whole network, so this is the tip of the iceberg. The BOM's "no further duplicates" claim is not a finding of absence; it is a statement that the search was not conducted at the block level. --- 4. Loss of Decimal Precision (1995–2011) The *Lost Decimals Report* documents 51 distinct loss windows across 23 ACORN-SAT stations, totalling 51,249 station-days where the fraction of zero-decimal (whole-number) readings was between 90% and 100%. - These windows occurred almost exclusively between 1995 and 2011 — the era when automated digital weather stations were introduced. - Examples: - Cape Otway Tmin: 3,533 consecutive days (~9.6 years) with 99.7% zero decimals. - Katanning Tmax & Tmin: Over 2,800 days with 98.9%–99.2% zero decimals. - Cabramurra Tmax & Tmin: 2,403 days with 99.7% zero decimals. - Camooweal Tmin: 1,676 days with 100.0% zero decimals. Why it matters: Whole-degree rounding systematically distorts variance, extremes, and percentile distributions (quantisation bias). It also creates artificial breakpoints in homogenisation algorithms when whole-integer values are fed into regression matrices alongside high-precision neighbour data. This is a modern-era data failure, not a historical quirk. --- 5. Machine-Learning Predictability and One-Way Adjustments - Sydney's Max T adjustments are 100% positive across 110 consecutive years (not 10 years, as some reporting has suggested). - Random Forest models show 70% of adjustment variance (R² ≈ 0.70) is predicted by calendar decade alone, with zero weather data input. Why it matters: Physical site changes (moves, screen replacements) should produce random, bi-directional adjustments. A 110-year one-way series that is predictable from calendar era alone is the signature of mathematical curve-fitting, not physical metadata correction. --- 6. Cross-Station Adjustment Propagation - Sydney's adjustments explain 71% of the adjustment variance at Bourke, NSW (R² = 0.711) — a station 800km away in a different climate zone. Why it matters: If a processing artifact at one station predicts adjustments at a distant station, the issue is systemic across the ACORN-SAT network, not local. --- 7. Structural Signatures: Flatlines, Value Traps, and Linear Regression Blocks - 17 consecutive days of exact 31.0°C at Palmerville (p ≈ 10⁻³⁴). - Over 800 value traps identified (repeated exact values trapping homogenisation algorithms). - 3+1+3, 4+1+1, and 5+1+5 structures (e.g., 5 days at 37.8°C, 1 day at 38.9°C, 5 days at 37.8°C) at a scale natural variation does not produce. - Hundreds of three-day linear-regression blocks between maximum and minimum temperatures. - ACORN itself creates flatlines: 10 days of flatline identical temperatures were created at Charters Towers from different incoming raw temps, and 13 days at Palmerville, with ACORN making a 6°C adjustment to one day to bring it in line with the flatliners. Why it matters: These are structural signatures — the shape that editing, infilling, and carry-forward operations leave in a dataset. Whether any individual case is natural is a question the BOM can answer for each one; the point is that the frequency and regularity of these structures across the network is not consistent with natural variation. --- Summary Table | Finding | Scale | Statistical Significance | | Whole-month replications (Sydney) | Jun 1943/1944; Nov 1903/1905 | Both flagged Quality = Y | | 1935/1936 seam | 14 days + 2-day insert + malformed value + 14 days | Block splice signature | | Duplicate block events | Over 1,000 blocks, sizes down to 5 days | p < 0.0001 (Monte Carlo) | | Post-1990 duplication | 42% of all events | Not consistent with pre-1946 explanation | | Decimal precision loss | 51,249 station-days across 23 stations | 90–100% zero-decimal readings | | One-way adjustments | 110 consecutive years (Sydney Max T) | Random Forest R² ≈ 0.70 from calendar decade | | Cross-station propagation | Sydney predicts Bourke (800km) | R² = 0.711 | | Flatlines | 17 consecutive days at 31.0°C | p ≈ 10⁻³⁴ | | Value traps | Over 800 | Structural signature | --- Conclusion The BOM's statement acknowledges 4 full-month duplicates and attributes all other patterns to "well-documented characteristics" or "misunderstandings." This does not account for: 1. The whole-month replications at Sydney, which fall inside the BOM's own pre-1946 window and are flagged Quality = Y. 2. The 1935/1936 seam, which shows the mechanism of block splicing. 3. The 42% post-1990 duplication rate, outside the pre-1946 explanation. 4. The modern-era precision loss (1995–2011). 5. The 110-year one-way adjustment series, predictable from calendar decade alone. 6. The cross-station propagation of adjustments across 800km. 7. The structural signatures (flatlines, value traps, linear regression blocks) that are not consistent with natural variation. The BOM's statement contains a specific claim — that no further duplicate months exist — that its own published data contradicts. Until these patterns are inventoried, flagged, and either confirmed as real or removed, the claim that the dataset is now clean is unsupported. --- Contact: Thomas Berger | tomberger2@gmail.com | 0425 267 098