Source materials: UNFCCC documents in Bracket

This statement covers the documents Bracket holds. It is written to be used three ways: in the app (on the text library page), in the footer of a report or paper, and as the provenance section of a public dataset should one ever be released. The short forms are at the end.

What the corpus is

Bracket holds a collection of public UNFCCC documents, assembled so that researchers can cite the exact wording a meeting was arguing about, at the version it stood at when they were in the room.

It contains:

• Adopted decisions and conference reports, including the addenda that carry the decisions.
• In-session documents: draft conclusions, draft decisions, L documents, CRPs, informal notes and non-papers, kept as a chain of successive versions rather than only the final text.
• Foundation texts: the Convention, the Kyoto Protocol, the Paris Agreement, the rules of procedure as applied.

As of September 2026 the collection register holds about 1,350 document versions, of which roughly 950 are published in the app's library, across COP29, COP30, SB62 and SB64 plus the foundation texts, in English only. The full index, one row per document version with its symbol, source, date and checksum, is docs/corpus/register.csv in the project repository. [Update the figures when this document is issued.]

Where the documents come from

Four sources, all public, none requiring a login:

Collection is deliberately slow: one file every 8 to 15 seconds, no credentials, no cookies, an identifying user agent, and the fetcher stops at the first response that is not a PDF. Pages on unfccc.int that sit behind bot protection are not read; nothing about that protection is circumvented.

Every fetch is recorded in a manifest (docs/corpus/*.manifest.json): the address requested, the time, the file size and a SHA-256 hash of the bytes received.

The documents are reproduced unaltered

The PDF we received is kept exactly as received and is the thing of record. Its SHA-256 hash is stored with it, so any copy can be checked against the original.

For reading and citation, the text is extracted from that PDF and split into paragraphs. Extraction does not change wording. Where extraction is imperfect (a table, a two-column layout, an automatically numbered list from a Word original), the fault is in the extraction, not in a decision to edit; docs/corpus/fidelity-2026-09-18.md reports how the extraction is measured against a hand-made reference set.

Paragraph references are the document's own. "¶4" means the paragraph the document numbers 4, not the fourth paragraph of our split. Preambular paragraphs are labelled pp1, pp2, pp3 as the room counts them. A citation in a note therefore points at the same place as a citation in the meeting.

The copyright position, and its conditions

The UNFCCC's terms of use state that official texts, data and documents on its site are in the public domain and may be freely downloaded, copied and printed, provided the content is unchanged and the source is acknowledged. The project relies on that for the documents themselves, and meets both conditions: unchanged content, source recorded per file.

A second clause pulls the other way. The same terms grant the site's content for personal, non-commercial use, expressly without the right to resell or redistribute it, or to compile or create derivative works from it. A compiled corpus with our structure on top of it is exactly what that clause names. The project's position, decided 19 September 2026, is therefore:

• Inside the team: collecting, holding and using the documents for research is within the public-domain grant and within personal, non-commercial use.
• Outside the team: before the corpus is published, shared or redistributed as a collection, written permission is sought from the secretariat ([webmaster@unfccc.int], Communications and Engagement Division), and party submissions and presentations are considered for exclusion, because they may not be "official texts" at all and their authors are not the secretariat.

This is a reading of published terms by researchers, not legal advice. No lawyer has confirmed it and none is engaged (names-review-2026-09.md, decision 12): the project proceeds on this reading inside the team, and asks the secretariat's permission before any publication of the collection (README.md, the non-commercial undertaking).

What the project undertakes. Bracket is an instrument and never a product; access to it is never sold; the corpus is never redistributed without permission; and a grant that pays a researcher to do research is not commercial use of these documents. The undertaking in full, with the rights statement, is in README.md in this folder.

What we add, and how it is marked

Our own work on top of the documents is a separate layer. It is always distinguishable from the source text in the interface, in an export and in any published dataset:

• Paragraph splitting and labelling, including the document's own reference numbers.
• Version chains: which draft superseded which, within one negotiation's paper trail.
• Topic and agenda-item codes, mapped to the project's own scheme.
• Cross-reference resolution: turning "decision 1/CMA.5, paragraph 28" into a link. References are resolved while a document is read and are not stored as claims about the text.
• Extraction metadata: page counts, paragraph counts, bracket and option counts.
• Researchers' citations: a link from a field note to a paragraph. These are research data, not part of the corpus, and follow the rules in privacy-notice.md.

In an export, source text and annotation are in different columns and different files, and the codebook says which is which. If the annotation layer is ever released, it is released under CC BY 4.0 (decided 24 September 2026) while the source documents keep their own terms, which is the usual arrangement for projects that redistribute official texts.

What is not included

• Lists of participants in the library. The UNFCCC's published lists of participants are held in the local archive since 25 September 2026, flagged, and never imported into the library, parsed for names, indexed or shown to members. Until then they were excluded from every run.
• Contact details in the library. No delegate's or secretariat official's contact details are imported into the library or shown in the app. The local archive of the Internet Archive's captures keeps UNFCCC landing pages whole, as published, and some of them name the secretariat contact for an agenda item; those pages are not parsed for names, not indexed and not shown to members. Whether they are redacted is a later, separate decision (25 September 2026).
• Third-party analysis. The Earth Negotiations Bulletin, Carbon Brief's own written coverage and similar work are not in the corpus. Where a third-party tracker was used to find documents, only the published counts and the addresses of UNFCCC files were taken, and their own text was not.
• Other language editions. English only in the library; other UN language editions stay in the local archive and are not imported.

Short forms

For a report footer:

Source documents are public UNFCCC texts, reproduced unaltered from docs.unfccc.int and public archives, with the symbol, version and source recorded for each. UNFCCC official documents are in the public domain where the content is unchanged and the source is acknowledged. Paragraph splitting, version chains and topic codes are added by Bracket.

For the app's library page:

Public UNFCCC documents, kept as published. Paragraph numbers are the document's own. Source and capture date are on every text. Not for redistribution outside the team.

For a citation in a paper:

UNFCCC, [Title], [SYMBOL] ([DATE]), paragraph [N].

Contact for questions about the corpus: contact@bracketresearch.org.