Dataset Release · De-identification and provenance
Release a dataset to an AI lab with a record both sides can verify.
You hold data a lab wants to license. Before it leaves your environment, direct identifiers are removed, opt-out lists are applied, and the sources and rights are written down. The result is a signed release record the buyer verifies offline, with a free open-source tool. The tools run where the data is; we never receive it.
Why now
The lab's paperwork now lands on the seller.
Providers of general-purpose AI models must publish a summary of their training content on the EU template, including licensed and private data sources. The AI Office can enforce this since 2 August 2026, so labs now ask the broker for the documentation.
Since the Court of Justice's ruling in EDPS v SRB (C-413/23 P, September 2025), whether pseudonymised data is personal data depends on what the recipient can realistically do with it. A release is a question about one recipient, not a property of the file.
California's Delete Act: since 1 August 2026, registered data brokers must process the state's DROP deletion list every 45 days. A dataset licensed for training needs to show which suppression list was applied, and when.
Training-data licences ask for a provenance file, warranties on rights and consents, and suppression of opted-out people. Someone has to produce that evidence, and it should not be a spreadsheet written after the fact.
The release record
What the buyer receives with the files.
- which direct identifiers were removed, pseudonymised or generalised, column by column, under which policy;
- how many values each rule changed, never the values themselves;
- which opt-out and deletion lists were applied, identified by their hash and date;
- the sources and rights you declare for the data, and the recipient the release is for;
- hashes of the source and of every released file, so the buyer can check that the files they hold are the ones the record describes.
The record is signed with your key. Each new release of the same dataset, for example after the next deletion cycle, links to the one before it, so the buyer can follow the history. The record format and its verifier are being built in the open; design partners shape the first versions.
The offer
From one workshop to a signed release.
-
Start here
Dataset Release Workshop
Half day remote or one day on your premises
- Your data dictionary and schema: direct identifiers, quasi-identifiers and free-text fields
- Your rights chain: sources, consents, opt-outs and acquisition history
- The buyer's asks mapped to what you can document: provenance file, warranties, suppression lists, training-content summary fields
- A written readiness note within three working days: blockers, a draft column policy, a fixed quote for the design
from €2,400 (remote) · €4,900 (on site)
Net prices, plus travel for on-site days
-
Then
Dataset Release Design
Per dataset, 3–4 weeks
- A column policy: drop, pseudonymise with your own key, generalise, scrub free text, or keep
- Key handling, suppression-list procedure and the re-release cycle
- A provenance file mapped to the Data Provenance Standards and Croissant fields
- A sources-and-rights schedule as drafting input for your counsel, and a recipient-risk checklist
from €12,000
Net price; the workshop fee is fully credited if you order within 60 days
-
Design partners
Release pilot
8–10 weeks, in your environment
- You run the toolset on one real dataset, in your environment
- We see the release record (hashes and counts), never the records
- One signed release record your buyer can verify
- Your requirements shape the first versions of the toolset
On request
Limited to a small number of partners while the toolset is in development
For business customers only. All prices are net; for business customers in other EU member states, VAT is accounted for by reverse charge.
The difference
The seal or the letter.
De-identification tools transform data inside your environment and stop there; what the buyer receives is a file and your word. We add the second half: a signed record of what was removed, what remained, which sources and rights, which suppression list and when. The seal proves the dataset the lab holds is the one the record describes; the letter proves who released it. The key stays with you, and the buyer verifies without an account and without trusting us.
- Status
- The dataset tooling is in development. It builds on shipped, open-source components: PII detection from presidio-hardened-x402, signed evidence from presidio-evidence, and the ingest ledger and differentially private training of presidio-hardened-ai.
- Where it runs
- In your environment, on your infrastructure. PRESIDIO never receives your dataset, and nothing is processed on a hosted service.
- What we need
- For the workshop and the design: schema, data dictionary and contracts. Synthetic sample rows at most, never real records.
- For the buyer
- The record format and the verifier are open source (MIT). A lab checks a release without an account and without trusting us.
Components: presidio-hardened-x402,
presidio-evidence,
presidio-hardened-ai.
How it works
What happens after you book.
- A 30-minute call: the dataset, the buyer, your timeline.
- A fixed-price quote the same day.
- A preparation list: schema, data dictionary and source contracts, two working days before the date. No data.
- The workshop, remote or on your premises.
- The readiness note and a quote for the design within three working days.
Questions
What sellers and buyers ask.
Is the released dataset anonymised afterwards?
The record does not say so, and neither do we. It documents what was removed, what remained and for whom. Whether data is personal for a given recipient is a legal question for you and your counsel; the record gives them the facts to answer it.
Do you replace a HIPAA Expert Determination?
No. We are not an expert of record. If your release needs an Expert Determination, the release record documents what the expert's method required and that it was carried out.
Do you add a watermark to the dataset?
No. Watermarks and embedded assertions are easily stripped. The release record binds the files by their hashes instead, and travels with the licence.
Will you see our data?
No. The tools run where the data is. For the workshop and the design we work from schemas, data dictionaries and contracts. Please do not attach data to an offer request.
We bought the data from a company that no longer exists.
Then the rights chain is the first thing we look at. Where it cannot be documented, we say so, and we do not take the release on. A record makes provenance visible; it does not clean it.
Who is behind PRESIDIO?
PRESIDIO OOD is a company based in Sofia, in the EU. Vladimir Stantchev leads every engagement personally, and the tools are open source, so you are never locked in.
Request an offer
A few details are enough.
Please describe the dataset; do not attach any data.
A release record documents what was done. It is not an anonymisation opinion, not a conformity assessment and not legal advice; whether data is personal for a given recipient is a legal question.