How Climate Errata works
An independent register of errors and issues in climate science: published papers, gridded and station datasets, and the software used to analyse them.
What we check
Methods, statistics, data processing and code. Most entries come from reproducing a result with the published data and code, or from comparing a dataset with independent references and station records.
Contact first
Every finding is sent privately to the authors, data producers or developers before anything is posted. An entry appears here only after they reply or three weeks pass without a reply. Their response is recorded on the entry.
Fixes we propose
We don't just report problems. Where we can, we propose a fix and send it with the report:
- Software: a patch or pull request, with a test that fails before the fix and passes after it.
- Datasets: the cells, regions and periods affected, the likely cause (for example a change in the station network), and a workaround for users, such as masking the affected period or checking against another product.
- Papers: a corrected analysis that shows how much the result changes, with the code to reproduce it.
These are recommendations. The authors, data producers or developers decide whether and how to fix their work, and each entry records what they did. The "What users can do" part of an entry says how to avoid the problem in the meantime.
Statuses
- Reported sent to the responsible party; no resolution yet.
- Acknowledged the responsible party agrees there is an issue.
- Disputed the responsible party disagrees; both views are shown.
- Fixed corrected in a new version, erratum or release. "Fixed – not yet released" means the fix is merged in the code but no release contains it yet. Each fixed entry gives the fix date, the version that contains it, the affected versions, and whether the fix is retroactive: a software fix changes results only from that version on, and results already computed with affected versions stay wrong until rerun; a dataset fix may or may not reprocess the past record.
- No error found checked; the claims or code hold.
- Already known the issue was already documented elsewhere.
How the work is done
The analyses were carried out with an AI system (Claude, Anthropic) under the author's direction. The author checked the methods and results and is responsible for the content.
Our reports on Zenodo
Write-ups of single findings are linked from their entries. Two broader reports cover the work as a whole:
- AI-Assisted Auditing of Climate Code and Statistical Methods: What Was Found, What Was Not, and Lessons for the Climate Community (November 2026).
- When Statistical Assumptions Go Unexamined: Common Methods, Their Assumptions, and Eight Influential Atmospheric-Science Cases (October 2026).
Significance
Each finding that claims a problem is scored 0–2 on three criteria. The class comes from the total (0–6): High 5–6, Medium 3–4, Low 0–2. Clean checks and already-known issues are not scored.
| Criterion | 0 | 1 | 2 |
|---|---|---|---|
| S · Size of the effect, relative to the signal users take from the output | < 5% of the signal, or inside stated uncertainty | 5–25%, or moves a result across a significance threshold | > 25%, reverses the sign or conclusion, or the output is invalid |
| R · Reach of the affected code, data or paper | Niche: non-default option, under ~1,000 downloads a month, or one paper with few citations | Moderate: standard tool in a sub-field, tens of citing papers, or a recent paper in a high-profile journal | Infrastructure: CMIP/IPCC/ISIMIP or national datasets, the dominant tool of its kind, 100+ citing papers, or policy and press use |
| C · Consequence for published or operational results | None shown | Affects results in principle (including bugs confirmed or fixed by developers) | A specific published headline or production dataset is shown to change, by our rerun or the authors' own statement |
Reliability of the scoring
Independent raters applied the written rubric blind to all 80 findings in the full audit (public and not yet public). Agreement with the first rater (weighted Cohen's kappa), rubric v1 → v2: Size 0.77 → 0.81; Reach 0.68 → 0.73; Consequence 0.62 → 0.73; Class 0.62 (65% exact) → 0.70 (79% exact). Between rounds one rule was clarified: a developer-confirmed or merged fix alone counts as C = 1, not C = 2. All raters so far are AI instances; a human rater on a random sample is planned. Reach scores marked "est." are estimates to be replaced by measured download or citation counts.
Rules
- We discuss methods and data, never people or motives.
- Every claim is backed by evidence that can be rerun.
- If we get something wrong, we correct the entry openly and date the change.
Policies
The full policies (author responses and corrections, moderation, code of conduct, funding and conflicts of interest, privacy, takedown requests) are on the Policies and privacy page.
Site updates
Changes to the website itself are listed on the Site updates page, with an RSS feed.
Contact
Write to us · GitHub klemmeri · Substack @markmorrissey3.
Discussion and questions
Climate Errata is also a place for researchers to ask about the datasets and papers they use. Every register entry has a discussion thread. Any paper or dataset can have one too: enter its DOI on the Discuss a DOI page, and everyone asking about the same work shares one thread. Ask other readers, the authors or the maintainer. Search covers both the register and the discussions.
Threads use GitHub Discussions, so commenting needs a free GitHub account. Without one, use Write to us.
Papers that use affected data
For a problem in a dataset, the entry may list papers whose abstract or text shows they use the affected data for the affected region and period. The list is a pointer for readers and authors, not a finding about those papers: the effect on each paper is not assessed, and many conclusions will hold. Papers whose authors we are in contact with are not listed until that exchange has had its three weeks. Authors can reply on the entry or write to us, and we correct the list on request.
Posting policy
- Questions, answers and discussion are welcome on any thread: how a dataset was built, what changed between versions, whether a method suits a purpose, how to reproduce a result.
- New error claims go to the responsible party first. If you think you have found an error that has not been reported, tell the authors or data producers, or send it to us privately. We check it, contact them and give them at least three weeks, the same rule we follow ourselves.
- Comments that make a new, unreported error claim about someone's work may be hidden until the responsible party has been contacted. They are restored once that has happened. Claims about issues already public (a published erratum, a known-issues page, an open GitHub issue) are fine; please link the source.
- Discuss methods and data, never people or motives. Be specific and give the version, region and period you looked at.
- Authors and data producers are invited to reply on any thread, or to send a response for the entry itself through Write to us.
- Off-topic, abusive or promotional comments are removed. Moderation decisions are made by the maintainer.
Reuse
Entries are licensed CC BY 4.0 and the register is available as entries.json. Code is MIT-licensed.