Data License Agreement
Version 2026-08-12. Effective 2026-08-12.
This licence says what you may do with the data we deliver. It is written to be unambiguous on the point that matters most to a team building retrieval systems: you may use our datasets as a retrieval corpus, you may generate embeddings and other derived representations from them, and you keep that right for the releases you received even after you stop paying.
It is part of your agreement with Andrii Sukhanov, trading as merkleset (see the Legal Notice), and works together with the Terms of Service. Where this licence and the Terms of Service differ about permitted use of data, this licence controls.
1. What is licensed
1.1 Licensed Data means every dataset release we made available to you while your subscription was in effect, including the JSONL data files, the per-record provenance fields, the release manifest and its Merkle root, and our derived and enriched fields.
1.2 A release covers the collection window it was produced for. Each release we make available to you during a paid term is Licensed Data, whether or not you download it while the term is running.
1.3 Licensed Data does not include our software, our pipeline, our internal raw evidence store, or any release published after your subscription ended.
2. Licence grant
2.1 We grant you a non-exclusive, worldwide, perpetual, non-transferable, royalty-free licence (after payment of the fees for the relevant term) to use, reproduce, store, modify and create derived works from the Licensed Data for your own business purposes and for the purposes of your products and services.
2.2 The grant expressly includes the right to:
- load the Licensed Data into a retrieval, search or retrieval-augmented generation system, as a corpus or an index;
- compute embeddings, vectors, indexes, summaries, classifications, knowledge graphs and other derived representations from it, and to keep and use those representations;
- fine-tune, train, evaluate and benchmark machine learning models on it;
- use the outputs of any such system or model, including outputs shown to your own end users, without any further fee or attribution obligation to us;
- use it for internal analytics, reporting and research;
- combine it with your own data and with third-party data you are entitled to use.
2.3 Sub-licence to your group and your contractors. You may extend the rights in clause 2.2 to your affiliates and to contractors and processors acting for you, provided they use the Licensed Data only for your purposes and you remain responsible for their compliance.
2.4 Users. Unless your order form states a different scope, the licence covers the organisation named on your account and the individuals it authorises. Credential sharing beyond that scope is prohibited by the Acceptable Use Policy.
3. What happens when your subscription ends
Two different things end at two different times, and conflating them is the most common worry about buying a corpus. So, plainly:
Access ends. The licence does not.
3.1 Access ends on the day your subscription ends. No further downloads, no new releases, and no new signed URLs. This is enforced technically by the entitlement check on every download request, not just promised here.
3.2 The licence survives. The licence in clause 2 for Licensed Data delivered or made available to you during a paid term is perpetual and irrevocable, and it survives expiry, cancellation or termination of your subscription for any reason other than clause 3.3. You are not required to delete, unwind or re-compute embeddings, indexes, fine-tuned models or other derived representations because your subscription ended. Cancelling stops new deliveries; it does not reach backwards into work you already built.
3.3 The only case in which the licence ends is termination by us for your material breach of this licence — for example redistribution of the datasets, or an attempt to re-identify individuals — and then only if we give you written notice and you do not cure the breach within 30 days where a cure is possible. Even then, outputs you already published in good faith to your own end users do not become infringing.
3.4 Perpetual licence is not perpetual access. Your right to download ends with your subscription. Keep your own copies of any release you rely on; see clause 5.4 of the Terms of Service.
4. What you may not do
4.1 No redistribution of the datasets. You may not sell, licence, publish, share or otherwise make available the Licensed Data, or any substantial part of it, as data — whether in our format or reformatted, and whether for a fee or free of charge.
4.2 No competing data product. You may not use the Licensed Data to build, train or populate a product whose purpose is to supply datasets, corpora or data feeds to third parties, and you may not sub-licence it as part of a data product.
4.3 No re-identification. You may not attempt to identify any individual from the Licensed Data, whether alone or by combining it with other data, and you may not use the Licensed Data to enrich a profile about an identified or identifiable individual.
4.4 No eligibility decisions about people. You may not use the Licensed Data to make or support a decision about an individual's eligibility for employment, credit, insurance, housing, or a similar benefit, and you may not use it in a consumer report.
4.5 No misrepresentation of provenance. You may not alter our provenance fields, manifests or hashes and then present them as ours, and you may not present modified data as verified by us.
4.6 What clause 4.1 does not restrict. It does not restrict: outputs of a model or a retrieval system that answer a user's question, including short quotations of source records in an answer; aggregate statistics, counts or metrics derived from the data; a small number of individual records shown or shared to illustrate a point, support a decision or satisfy an audit; or your own re-collection of the same public records from the original government sources, which you are always free to do.
5. Attribution — permitted, not required
5.1 You do not have to credit us. There is no attribution condition on the outputs of your system, and none on your product.
5.2 If you want to credit us, this form is accurate and we welcome it:
Source data: merkleset — {dataset name}, release {version}, Merkle root {root}. https://merkleset.com
5.3 Do not describe your product as certified, verified or approved by us. Stating that you use merkleset data, and stating a release version and Merkle root, is accurate and permitted.
6. The underlying source material
6.1 Our shipping datasets are built from official public United States federal government sources. Works of the US government are not subject to copyright in the United States, and we claim no copyright in that source material.
6.2 What we licence to you is our own contribution: the collection, the normalisation and cleaning, the chunking, the derived and enriched fields, the provenance layer, and the selection and arrangement of records into a versioned dataset. That contribution is what clause 4.1 protects.
6.3 You may always obtain the same underlying records from their government sources yourself. This licence does not, and could not, stop you.
6.4 If we ever add a dataset whose source material is not a government work, we will state its licensing terms on that dataset's page before you can subscribe to it, and those terms will govern that dataset.
7. Personal data in the datasets
7.1 We build datasets to contain no personal data, and we enforce that in the pipeline rather than by policy alone. Screening runs before any published hash is computed, so the hashes describe exactly the bytes you receive. Email addresses and telephone numbers are redacted. Values shaped like a US Social Security number with an explicit nearby cue cause the whole record to be dropped. Person names are redacted where a source field declares them or where an explicit contact label introduces them. For the procurement corpus we go further and never read the source's contact columns at all.
7.2 The scope of that screening is deliberately narrow, and we publish its gaps. It does not catch a person named in running prose without a contact label, labels separated by whitespace only, postal addresses, obfuscated contact details, or number shapes with no nearby cue. Those exclusions are chosen so that dates, docket numbers, classification codes and numeric tables survive intact.
7.3 So the accurate statement, and the one you may rely on, is that the datasets are systematically screened to a documented and deliberately narrow scope. We do not warrant that a dataset contains no personal data.
7.4 If you find personal data in a release, tell us at privacy@merkleset.com. We will remove it from future releases and delete it from our internal evidence store. We cannot recall a release you already downloaded, so you must remove the identified data from your own copies and indexes on our reasonable request, and we will tell you exactly which records are affected.
7.5 You must not use any residual personal data you encounter. Clauses 4.3 and 4.4 apply to it in full.
8. Verification
8.1 Every release ships a manifest with a Merkle root over the hashes of the chunk lines,
plus a hash and byte count for each file. You can recompute all of it from the files you
downloaded, with no tool of ours: hash each line of chunks.jsonl, sort the hashes, build
the tree, and compare the root against the manifest.
8.2 We publish per-record provenance so that you can trace any record back to the source URL, the fetch time, the HTTP status and the hash of the raw source bytes we retained.
8.3 The raw source snapshots behind snapshot_ref are held internally and are not delivered
to you. They are unscreened, which is exactly why they stay internal. snapshot_ref is an
internal reference, not a URL you can fetch.
9. Warranty and liability
9.1 We warrant that we have the right to grant this licence, and that we collected the source material in the manner described in clause 6.2 of the Terms of Service.
9.2 Apart from clause 9.1, the Licensed Data is provided without warranty. The disclaimer in clause 9.2 of the Terms of Service and the liability limits in clause 10 of the Terms of Service apply to this licence, including to any claim about the content of the Licensed Data.
9.3 The indemnities in clause 11 of the Terms of Service apply to this licence.
10. General
10.1 This licence is versioned by effective date. A new version applies to releases delivered after you accept it; it does not change the licence attached to a release already delivered.
10.2 English is the authoritative language of this licence; a translation is for convenience only.
10.3 The governing law and venue are those in clause 15 of the Terms of Service: the laws of the Kingdom of Spain, and the exclusive jurisdiction of the courts of Santa Cruz de Tenerife (Canary Islands, Spain).
10.4 Questions about scope, including whether a specific use is permitted, go to legal@merkleset.com. We would rather answer than have you guess.