How every number is sourced and checked

Where the numbers come from

Each number comes from an organisation that measures or compiles it: national science agencies, international bodies, research groups and peer-reviewed papers. We use the producer's own files rather than copies on other sites. Where a widely used aggregator processes the same data, we still cite the producer. Every source is listed on Sources with its citation, licence and current status.

One series comes from one source. We never stitch two sources into one line. Where respected sources disagree, we show both and explain why.

How a number is fetched and kept

1. A scheduled job downloads each file from the producer's address and records the date, the server's version headers, and a sha256 fingerprint of the exact bytes. If the file is unchanged, nothing is rebuilt.
2. The bytes are stored in our archive. Where the licence allows, that copy is public, and the producer's page is also captured by the Internet Archive's Wayback Machine, so a number stays traceable even if the original moves or disappears.
3. Some producers block automated downloads. Their files are downloaded by a person and recorded the same way, and the data page says so.

What we do to the data

As little as possible, and every step is written down. Each number's data page lists the steps in plain words, such as “re-expressed relative to the dataset's own 1850–1900 average”, along with the code that did it and a fingerprint of that code. Values are stored unrounded and rounded only for display. Nothing is filled in: a gap in the data stays a gap.

• Temperatures are shown relative to 1850–1900 where the dataset covers that period, using that dataset's own average. Where it does not, the data page states the baseline used.
• Greenhouse gas totals state which warming factor they use to combine gases (for example 100-year global warming potentials from the IPCC's fifth or sixth assessment) and whether land use is included.
• Preliminary values, such as the current month or a year in progress, are labelled as such.

How numbers are checked

Where a producer states a value in its own words (for example in a press release or report), an automated test compares our number with that statement for the same version of the data. Each published number is also re-derived independently from the stored file before it first appears on the site. Sentences that claim a record or a ranking are checked against the data every time it updates, so an update cannot silently make the text wrong.

Licences

Each producer sets the terms for its data, and the site follows them. Most data can be shown and shared with credit; you can download those numbers from their data page under the same terms. Some producers allow sharing only for non-commercial use, or ask that derived versions carry the same licence; the download says so. A few allow their published figures to be shown but not redistributed or recalculated. Those numbers appear with a citation and a link to the original, without a download. The licence of every figure is shown beside it.

Updates and corrections

The data is checked weekly. When a producer publishes a new version, the affected numbers update and their data pages record the new version. If a source fails to update or changes format, its numbers stay at the last version that passed our checks, and Status says so. Our own mistakes are fixed and listed on Corrections.

Reproducing the numbers

The pipeline that fetches, transforms and checks the data is public at GitHub. Its output for each number lists the fingerprints of its input files, so you can confirm you are working from the same bytes.