Climpact Open Data
Traceable Reception and Attribution in Climate-law Evidence
Every reading says how it was made and can be checked. The research team judges the citations in this release: it confirms those that cite the judgment, correcting them where needed, and discards the others. The other readings are proposed by the machine. TRACE 0.1 is the first release. How the proposals are made
Following climate judgments into the litigation that comes after them
For a climate judgment, TRACE finds the later decisions and filings that cite it, proposes what each citation does with it, and follows the argument onto the pages around it where the judgment is no longer named. Each proposal is labelled with the version of the method that made it.
What TRACE is for. TRACE brings to the surface what no research team could find by reading alone: every citation of the climate judgments it follows, across hundreds of later filings and thousands of pages, each with the sentence that carries it and a proposal of what it does, and hands it to the people who decide. People stay at its centre. Its strength is the pipeline. It stacks many layers on top of one another: name searches, paragraph maps, rules, five independent readings, models trained on the team's judgments, and a version code on every step. Each layer is measured, and every proposal can be traced back to the exact words on the page and the exact version that produced it. That order is what made these results possible with just over 400 citations judged by people. This first release is a beta, and a starting point, not a ceiling. Every new human judgment enters a pipeline that is already ordered and traceable to the smallest piece, and becomes clean training data, so each one makes the next model better informed. TRACE 0.2 will measure how far that takes it.
Presented at the Climpact International Inaugural Conference, 12 October 2026
Computed on ReCaS-Bari, the public data centre of the University of Bari and INFN01The map
Each citation TRACE has found, drawn on a map. Tap a document to read its citations, or follow one principle from court to court. Open Numbers for the influence of each judgment, with the TRACE Judicial Impact Index.
02Judicial impact
One number for how far each judgment has been taken up by the courts and the parties after it, built from every citation TRACE has read, with the range of what it could be.
Future Generations
Urgenda
Neubauer
KlimaSeniorinnen
Advisory Opinion OC-32/25
ICJ Advisory Opinion
1. What it counts
Every later case that cites the judgment counts once, however many of its filings name it. Each case is placed on the highest of four steps its citations reach, from a passing mention to a court relying on the judgment in its reasoning. The index is the sum of the weights of those cases.
2. Why these weights
We do not pick these numbers. All we assume is the order of the steps: a court relying on a judgment counts at least as much as a court naming it, and so on down. Averaged over every set of weights that keeps that order, the four weights come out as ¼, ½, ¾ and 1.
3. What the range means
The range is where the index lies once three doubts are counted together: that the machine misread some citations (estimated on the citations the team has checked), that the citing cases are a sample, and that the weights could be any of the admissible ones. It is computed a thousand times; the range holds nine in ten of the results.
4. What it does not do
It sums up reach, the parties' use and the courts' reliance; it does not replace them, because two judgments with the same index can be received in very different ways. It does not rank the judgments, which differ in age and in how far each search has gone. And it shows use, not cause: a court relying on a judgment does not mean the judgment decided the case.
Coming in later versions
Possible because TRACE holds granular data on every citation, proposed by the model through the TRACE pipeline and classified by people. Rule N01, beta.
03Numbers
The same numbers as the map: judgments, years, countries, courts, who speaks, how each citation is used, where it sits, the principles and the paragraphs. Tap any bar, row or country to filter, and every card recounts.
04
BetaThe citations
Search the sentences that cite the climate judgments, filter them by court, country, principle or type of citation, and open each one on the page of the filing where it stands.
The research team is verifying every citation, one by one. Until a citation is verified, it is shown as proposed by the machine, and you can help check it.
05What TRACE is
Later cases cite a judgment, argue with it and borrow its reasoning. TRACE records two kinds of uptake: citations that name the judgment, and reasoning that is taken over without naming it. It works for any climate judgment that has a case file.
A plain text search, with no AI, finds the pages that name the cited judgment in any of its forms. Each hit becomes a citation with a permanent code. The next question is what the citing text does with the judgment: it may only list it among authorities, describe it, rely on it, distinguish it or criticise it.
The name is searched in the text; DeepSeek V4 Flash, a large language model, proposes; the research team verifies, citation by citation.
A decision that cites a judgment often keeps using its reasoning on the pages around the citation without naming it again. Stage 1, in TRACE 0.1: follow the reasoning on up to ten pages before and after an explicit citation, in the same document. Stage 2, still to come: search the whole corpus for decisions that take up the reasoning without citing the judgment at all, with models trained on the team’s judgments from stage 1.
Stage 1: DeepSeek V4 Flash proposes the pages that continue the argument, and the team confirms them. Stage 2 comes later.
06What TRACE looks for
For every citation TRACE records two things: which principle of the judgment the citing text takes up, and what type of citation it is. Both come from written lists: the principles from a case file written for each judgment, the types from a codebook, the coding manual the research team follows.
Principles
A principle is something the judgment holds that other sources do not. Each is written as a short statement in English, next to the paragraph it comes from. A citation takes up a principle when the citing text relies on that statement, not merely when it names the case. One citation can take up several principles, or none.
Paragraph references: HR, Supreme Court of the Netherlands, 2019; RB, District Court of The Hague, 2015; Rn., marginal number of the Federal Constitutional Court; for Future Generations, the section of the judgment and its page. Urgenda’s five are shown in full in section 12, with the passages used to write them and the formulations left out.
URGUrgenda Foundation v. State of the NetherlandsSupreme Court of the Netherlands, 20 December 2019; District Court of The Hague, 24 June 2015
HR 5.7.7Own share: a small contribution is no excuse
Each country is responsible for its own share of global emissions and cannot escape that responsibility by arguing that its emissions are relatively small and that reducing them makes little difference on a global scale.
HR 5.6.2Articles 2 and 8 ECHR against a deferred, diffuse climate risk
Articles 2 and 8 ECHR oblige the State to take appropriate measures against the real and immediate risk of dangerous climate change, although it materialises only decades from now and threatens the population at large.
RB 4.83Duty of care: insufficient mitigation is unlawful
By failing to take sufficient mitigation measures against the high risk of dangerous climate change, the State falls short of the standard of due care owed to its citizens and acts negligently and therefore unlawfully.
HR 8.2.6An order to reach a result is not an order to legislate
A court may order the public body to take measures to achieve a certain goal, as long as that order does not amount to an order to create legislation with a particular content.
HR 6.4The court may set a minimum from scientific and international consensus
A court may require the State to achieve a specified minimum reduction of emissions by a given date, derived from the consensus of climate science and the international community on the State’s minimum fair share.
Types of citation
Every citation gets one type, from the codebook, which is final for TRACE 0.1 and will change in 0.2. The type says what the citing text does with the judgment. Who speaks and where the reference sits are recorded separately: a substantive citation by a party does not weigh the same as one by the court, nor a footnote the same as the body text.
01
Codebook labelmenzione_in_catena
The judgment appears in a list of precedents and the text draws nothing from it. It does no argumentative work: removing it would not change the reasoning.
Example, decided by a person
Dubetska and others v. Ukraine, App No 30499/03, 10 February 2011, §§ 145, 155. See also Urgenda Judgment (supra n.99), § 5.3.3. See also Tajkin v Turkey (supra n.86), § 119 and Hatton v UK (supra n.163), § 128 with further references.
URG-0421-JGreenpeace Nordic and Others v. Norway, petition, 2021, p. 43A party, in a footnote
The examples are single citations shown for their type, with their permanent code; they are not findings (section 14).
Who speaks
Recorded for every citation, whatever its type. The codebook names these roles and does not define them.
Where the reference sits
Read from the PDF, from the size of the type and its place on the page, and corrected by the coder only where it is wrong.
07Method
Seven steps for every cited judgment. Where a fixed rule is enough, a rule does the work. Where a text has to be read, a language model, DeepSeek V4 Flash, proposes. People then check the proposals. The research team judges the citations in this release: it confirms those that cite the judgment, correcting them where needed, and discards the others. The other readings are proposed by the machine.
01
A rule, no model
Plain text search, no AI
What it does
A plain text search looks for the name of the cited judgment on every page of the Climate Litigation Database, in all its forms: names, ECLI numbers, docket numbers, translations, other scripts and common OCR errors (OCR is text recognition on scanned pages). Pages whose text cannot be read are read again with OCR, but only if the PDF is the file the text came from, checked by its SHA-256 fingerprint. Each hit becomes a citation with a permanent code, such as URG-0146-J (a citation of Urgenda), NEU-0412-J (of Neubauer) or FGC-0033-J (of Future Generations). The last letter names the corpus: J for court decisions and filings, N for laws and policies, M for media. The code without the letter (URG-0146) still opens the same citation.
Versions in the registry
08Implicit uptake
TRACE looks for implicit uptake where it is most likely: on the pages around an explicit citation, in the same document, up to ten pages away. The ranking model scores the pages in that window; a fixed rule chooses the few worth reading; DeepSeek V4 Flash reads them with the citation and says whether the argument goes on without the name.
The name is searched in the text; DeepSeek V4 Flash, a large language model, proposes; the research team verifies, citation by citation.
Stage 1: DeepSeek V4 Flash proposes the pages that continue the argument, and the team confirms them. Stage 2 comes later.
09Codes and provenance
A citation code says which citation. A version code says how a proposal about it was made. The registry says exactly what that version was, down to the fingerprint of the model files.
Every version of the method that has produced a proposal is recorded here, with the model, settings, instructions and code it used, the fingerprints of its files, and what it changed from the version before. Open a version to see what it does, why it matters for the data and what replaced it. Where an instruction text was rebuilt after the run, the entry says so.
As of 19 October 2026: 61 versions in 7 families
Each tile is a date of the registry, the number in the corner how many versions it has. Select one to light up its versions in both columns.
10Public infrastructure
The reading of the citations and the training of TRACE’s own models run on ReCaS-Bari, the data centre built by the University of Bari Aldo Moro and INFN, the Italian National Institute for Nuclear Physics, on NVIDIA H100 GPUs. The Ask TRACE chat and the tools that read your own words are the exception: the chat’s answers are written by Google Gemini from TRACE’s data and texts, and Test a hypothesis, Your argument and Deep research send what you write to an external language model; see the privacy policy.
What runs there
Why it matters
TRACE 0.1 used 1,355 GPU-hours on NVIDIA H100 GPUs at ReCaS-Bari (25 September to 9 October 2026), about 0.8 to 1.5 MWh at the level of the data centre on a literature range of PUE, since ReCaS does not publish its own; the technical report gives the method and its limits.
11People and training
A proposal becomes a verified citation only when a person confirms it. What people decide is kept with its context, and it is what the project’s own models are trained and tested on.
Every citation
12Case files
For each cited judgment, a case file lists the principles it holds, each written as a short statement next to the paragraph it comes from. The paragraph map, the ranking model and DeepSeek V4 Flash all work from this list. Urgenda’s five are shown here, with the passages used to write them.
Principle
Each country is responsible for its own share of global emissions and cannot escape that responsibility by arguing that its emissions are relatively small and that reducing them makes little difference on a global scale.
Alternative formulation
A state's partial responsibility for a global harm does not lapse because others fail to act or because its own contribution to worldwide emissions is minor; each state answers for its part.
Source paragraph
Nor can the assertion that a country’s own share in global greenhouse gas emissions is very small and that reducing emissions from one’s own territory makes little difference on a global scale, be accepted as a defence. Indeed, acceptance of these defences would mean that a country could easily evade its partial responsibility by pointing out other countries or its own small share.
13Lab notebook
TRACE 0.1 comes out of a series of experiments. This is the record of what was tried, why, and what was kept, paused or set aside. It gives no scores, because the judgments they were measured on are being corrected; every comparison will be run again on the corrected data.
Search models that compare meaning (multilingual-e5-large, Qwen3-Embedding, Harrier, Nemotron, Perplexity pplx-embed) searched every page for the principles of the judgment, versions retrained on the team’s judgments ordered the candidates, and the team read them from the top.
Among the candidates, real implicit uptakes were rare: most pages shared the topic, not the source. Implicit uptake is now followed from an explicit citation first. The search across the corpus comes back after this first stage, with models trained on the team’s judgments; that search and its experiments are the starting point of the second stage on the judgments, and of the later work on laws and policies.
A second reading of the team’s early judgments, row by row.
The judgments had recorded whether a page was relevant to the judgment, not whether it took up its reasoning. The labels are being corrected, the codebook now separates the two questions, and every judgment keeps what the person was shown.
Reading each explicit citation with the pages around it, up to ten before and after, and asking on which of them the argument is taken up without the name.
The document has already chosen the judgment as a source, so the question stays on provenance. This is how TRACE 0.1 looks for implicit uptake.
Citations that give a paragraph number of the judgment, decoded by a hand-written map from paragraphs to principles.
Where a rule can answer, no model is used. The paragraph numbers were checked against the Dutch original of the judgment, paragraph by paragraph.
Untrained and trained rerankers of different families and sizes: bge, zerank-2, Qwen3-Reranker-4B and 8B, with LoRA adapters.
The 4-billion-parameter model was kept because it runs on modest hardware and its training can be published as a small file on top of the base model. It was trained on the early judgments, so it will be retrained on confirmed principles.
DeepSeek V4 Flash, a large language model with open weights, as a reader of citations and of their surrounding pages, with a briefing on the case and written coding rules, run with vLLM on our own GPUs at ReCaS-Bari: two readings for each citation at first, five with a majority vote since 2 October 2026 (rule R11, proposals by majority vote of five readings).
DeepSeek works from written rules and a briefing on the case rather than from examples. Its disagreements with coders on footnotes led to a new codebook rule.
Qwen3-14B with a LoRA adapter, trained on the readings of Urgenda by DeepSeek V4 Flash (distillation from DeepSeek V4 Flash), so that the work runs on a single GPU.
It matched DeepSeek V4 Flash on the type of citation, on citations that had been judged with DeepSeek’s proposals on the screen. It stays experimental: the proposals on the platform still come from DeepSeek V4 Flash until it matches it on citations judged without any proposal.
Variants of the search by meaning were first compared on how well they separated candidates already found.
That comparison rewards pushing down pages on the same topic, and real uptakes can go down with them. A first stage must above all lose nothing. The principle stays for the later work of that search.
Explicit citations of climate decisions, with the case name masked, set against pages that reason from the Paris Agreement, the UNFCCC or Articles 2 and 8 ECHR without naming any cited case.
Pages that make the same argument from another source are the right kind of negative example: they teach the model the difference between the same argument and the same source, with no human judgment needed for the cited case.
OCR of pages whose text came out as broken characters, in Korean, traditional Chinese and Spanish documents. In the first pass two documents were read from another PDF in the same folder, one of them with the same number of pages.
Text from OCR enters the corpus only if the PDF has the SHA-256 fingerprint recorded with the text, the same name and the same number of pages, and if the new text resembles the old one. The two wrong files are refused; the seven right ones pass. Reading the new text also showed that the name search missed citations written in Japanese forms of the case names, or split by a line break: it was corrected and run again.
Three decision models (Clef, Lux and Vega) that read after DeepSeek V4 Flash and before the research team, under a rule of adoption written before any number was known: enter only if the F1 on the principles rises above the proposal shown on the platform, with the interval of the difference above zero.
None enters: each stays below the proposal shown on the platform, on the principles. As a checker of the type of citation Vega catches 7 of the 18 wrong types, on too few rows to show an advantage; it is not used in this release. The proposals stay the five readings of DeepSeek V4 Flash.
Giving DeepSeek V4 Flash the most similar rows already judged by the team as precedents; and asking an open vision model, shown only the image of a page, whether the page cites the judgment.
Precedents bring no gain, on the type of citation or on the principles. The image found 132 of 136 pages of confirmed citations and said no on 40 of 40 pages that cite none of the judgments TRACE follows: a way to find citations where the text of a page is broken, not part of this release.
Reading many citations of a decision together, to propose patterns that can only be seen across citations, with Claude Opus 5.5 (Anthropic, through its API) as the instrument: a tool that proposes, not the author or the reviewer of the work. Every claim is tied to citation codes and checked automatically against the text.
It contributes no citation and no figure to TRACE 0.1. A sample of claims was reread when the reports were prepared, and the research team has not yet reviewed them: the findings are hypotheses for the process-tracing work of CLIMPACT, not results.
A second route to uptake without the name: passages of later filings paired with the paragraphs of a judgment, with a small open model discarding the pairs that are clearly not an uptake, calibrated on pairs with a known answer.
It discarded 21 of 88 known uptakes and turned 16 of 163 pairs discarded by the team into proposals: it did not meet the criterion written before the calibration, and its next step was not started.
14Honest limits
What TRACE finds are traces of uptake identified by the method, not proven influence. A trace says that a later decision or filing takes up reasoning matching the judgment; it does not say the judgment caused it.
15What comes next
Urgenda, Neubauer, Future Generations, KlimaSeniorinnen, the ICJ advisory opinion on climate change and the Inter-American Court advisory opinion OC-32/25, followed through the decisions and filings of courts and bodies worldwide. TRACE 0.2 will read the whole corpus again with our own reasoning reranker, trained on the team’s judgments and released open with TRACE 0.1.
The forms of its name, its principles and a map of its paragraphs are enough to follow a new judgment. New judgments are added release by release, including those other researchers ask for.
The same method will follow how judgments and their principles reach legislation and policy, starting with Climate Change Laws of the World (Grantham Research Institute with Climate Policy Radar), and then the other sources where climate-law reasoning travels.
A second stage will find where a judgment’s reasoning is taken up without the judgment being named, using models trained on the research team’s judgments.
A model trained on the team’s judgments and then frozen, so that anyone can rerun TRACE with the same model and the same instructions.
Agreement figures computed on citations the team judged without seeing any proposal of the machine.
TRACE will run on research requests: other judgments, other corpora, your own collections of documents. You can send a request below, under Use TRACE.
16Code
Everything behind the proposals is in one repository: the pipeline, the case files, the codebook and the registry of versions, with a copy of the instructions and the code of every version, and a note wherever a text was rebuilt after the run.
Code repository
TRACE
On GitHub
The pipeline: search, reading, paragraph map, ranking scores, putting the proposals together, codes, registry, training and testing of TRACE-Reader.
For each cited case: the forms of its name and its principles, the case briefing and the paragraph map.
The coding categories and the rules for hard cases.
The registry of versions, and the copies of instructions and code, with reconstructions declared.
On 3 October 2026 the code went through an independent external review, in three rounds. The problems it found were fixed; the full test suite now passes (509 tests, one skipped) and the maintained package reproduces the reference sample file for file (298 of 298). On the last round the reviewer asked for no further change beyond two small ones, and concluded that the package is sound. The review and our answers are in the repository.
Code comments are in English. The instructions given to the models, the case files and the codebook are in Italian, the working language of the team, and are published as they were run, except where the registry marks a text as rebuilt afterwards. The repository holds no document texts: what is published from the corpus is described under Source.
17Models
TRACE uses one model of ours in production (TRACE-Reranker 0.1), two experimental models of ours (TRACE-Reader and TRACE-Reranker that reasons) and DeepSeek V4 Flash, a model we did not train, which makes the proposals today.
Qwen3-Reranker-4B + LoRA
Scores a citation’s passage, and the pages around it, against each principle of the cited judgment. A rule then chooses the nearby pages worth reading.
Base model licence: Apache-2.0
Published with the Zenodo record
Qwen3-14B + LoRA
A smaller reader, trained on DeepSeek V4 Flash’s readings, that runs on one GPU. On the type of citation it agreed with the team about as often as DeepSeek V4 Flash did (0.60 against 0.57, on Urgenda, measured on 27 September 2026), but those judgments were made with DeepSeek’s proposals on the screen. The proposals on the platform still come from DeepSeek V4 Flash.
Base model licence: Apache-2.0
Published with the Zenodo record
Qwen3.5-9B + LoRA
Reads each citation, reasons in writing and proposes the principles of the cited judgment with the words that support them. Released as experimental: it was not used to produce the TRACE 0.1 citations. Trained several times from different random starts, it is on a par with the untrained model.
Licence: MIT (base model: Apache-2.0)
Published with the Zenodo record
Large language model, open weights
The large language model that reads every citation in context and proposes what it does, with a briefing on the cited case and the codebook: the proposals on the platform come from it. TRACE-Reader is distilled from its readings.
Model licence: MIT
Perplexity pplx-embed-v1-4b + LoRA
Paused, and not part of this release: it has no version code, no copy of its code and no fingerprint of its model files in the registry. The search by meaning that it served is set aside; it returns as the second stage on the judgments, retrained on the team’s judgments, and then on laws and policies (Climate Change Laws of the World).
Released with the data, in trace-0.1-team-judgments.csv: for every citation the team has judged, discarded ones included, the machine proposal next to the team decision, whether it was judged blind, and the second coder where a row was double-coded. Coders appear as pseudonyms; time spent and free-text notes are not published.
18Use TRACE
You can already ask TRACE questions about its data, download the citations and the corpus index and read the code. Soon CLIMPACT will have computing power of its own to run TRACE for other researchers: tell us now what you need.
Write what you are looking for as a sentence, in any language. TRACE turns it into the filters of the database and adds the citations closest in meaning to your words.
Open the databaseWhat you get
For example
German courts that distinguish Neubauer after 2022
Soon
Requests like these will run on machines of our own. Ask now: we answer every request and tell you what we can do, and when.
19Source
The Climate Litigation Database, read through its public API: a copy of 26 July 2026, with the updates of 19 September and October 2026. Data provided by the Sabin Center for Climate Change Law. Platform operated by Climate Policy Radar.
TRACE does not republish the documents. It publishes an index of their 578,146 pages in three tiers: an open index without text, the open text of public decisions, and the full index for researchers who ask.
Attribution
Data provided by the Sabin Center for Climate Change Law. Platform operated by Climate Policy Radar.
Court decisions and other official acts are not protected by copyright. The parties’ filings can be, and the database as a whole is a protected collection. So the text of pages is public only for official acts.
1Public, CC BY-NC 4.0
For each of the 17,039 documents and 578,146 pages TRACE read: identifiers, positions, language, whether the page was read by OCR and whether it is readable, and a SHA-256 fingerprint of the page text; the fingerprints of the PDFs; and where each citation found by the name search sits. With it comes a script that downloads the documents again from the database and checks them byte for byte against the fingerprints. It works for the 560,524 pages whose text comes from the PDF itself; the 17,622 pages read by OCR can only be checked in tier 3.
2Public
The text of every page of 4,275 documents issued by public authorities (judgments, orders, opinions and other official acts): 104,117 pages and the 1,004 citations of the dataset found in them. Left out as a precaution: 199 other documents of public authorities, those that carry the mark of a commercial legal database (61), translations (27) and decisions of United Kingdom courts (122); some fall in more than one group.
3Restricted, for research
The index with the text of every page, including the parties’ filings, for researchers who need to check TRACE’s results. It can be requested either by writing to trace@climpact.it or through the Zenodo record; access is granted by the CLIMPACT team after a request that states the research purpose and the institution. It is for that checking only: the file may not be redistributed, must be kept securely and deleted after use, and the source must be cited.
To ask for tier 3, write to trace@climpact.it or make the request through the Zenodo record, with your name, institution and research purpose. Request the full index
Each version stays available, so an analysis built on it remains reproducible.
Naglieri, G., Buono, E., Garofalo, C., Saracino, C., Dell'Atti, L., Rojas Verdugo, G. A. (2026). CLIMPACT TRACE 0.1: Traceable Reception and Attribution in Climate-law Evidence. CLIMPACT, University of Bari. Zenodo. https://doi.org/10.5281/zenodo.23275399
@software{naglieri_trace_2026,
author = {Naglieri, G. and Buono, E. and Garofalo, C. and Saracino, C. and Dell'Atti, L. and Rojas Verdugo, G. A.},
title = {CLIMPACT TRACE 0.1: Traceable Reception and Attribution in Climate-law Evidence},
year = {2026},
version = {0.1},
publisher = {CLIMPACT, University of Bari},
doi = {10.5281/zenodo.23275399},
url = {https://www.climpact.it/site/data/trace}
}20Credits
Authors of TRACE 0.1
TRACE 0.1