Sign up
Two sides, one record. A seller receives a receipt, a certificate and a statement. A buyer receives a package it can verify. Pick your side.
Seven short steps. Your engagement opens when you submit.
Start the supplier onboarding I'm buying dataTell us who you are and what you train. You see the catalogue once approved.
Start the buyer applicationAlready signed up? Seller portal or buyer portal: sign in with the email you gave, a one-time PIN, no password.
How it works, and what you can check
-
Deposit, with a custody receipt
receipt.json the seller receives it
You deposit into an isolated enclave through an upload link that can add and remove your files but never open them. Every file is hashed and the hashes are folded into one Merkle root. The custody receipt carries that root, the sizes and timestamps, and the signatures of two named custody officers. You can recompute the root from your own copy and compare.
Check it: Recompute the Merkle root from your own copy with intake verify; check both signatures against the published key ids.
The receipt commits to what was received. It says nothing about content, quality or rights.
-
De-identification behind a fail-closed gate
gate log the seller receives it
Inside the enclave, names, emails, secrets, identifiers and the git history itself are replaced by keyed stand-ins, the same stand-in for the same entity everywhere, so the links between a ticket, a commit and a conversation survive. The release is held by gates: a deterministic scan of every output byte for every value detected in the input, a check that planted canaries were all caught where canaries were planted, a secrets scan, a history rewrite check and structure checks. A single survivor blocks the release. The gate log records every result.
Check it: Read the gate log: six gates, each with its result, and the counts the run produced.
A passed gate is a measurement on this run, not a statement that nothing personal remains.
-
Packaging to the open standard, with a signed certificate
deidentification-certificate.json the seller receives it
The package follows the open standard: a manifest with every file hash and a Merkle root, the records in the shapes labs already ingest, a link table across tools, a rights file, a provenance manifest and a dataset card. The de-identification certificate pins the package by its hashes, states the method, the counts and the benchmark figures on a named sample, and lists what it does not claim. Two people sign it.
Check it: Ask for the reference validator and check the package offline: hashes, schemas, cross-references, the certificate and both signatures.
The certificate never says the data is free of personal data or that a legal standard is met. The standard rejects a certificate whose text makes either claim.
-
Evaluation without custody
dataset card, watermarked preview the buyer receives it
A buyer sees the dataset card, with statistics suppressed below a minimum count, and can ask for a preview: a small sample, watermarked to that buyer, with canaries that attribute any leak. The terms ride on every response. No raw deposit ever reaches a buyer, and no sample may be used to train a model.
Check it: Compare the manifest hash and Merkle root on the card with any delivery; check the preview file against its published SHA-256.
A preview is a sample, not a dataset.
-
Delivery in the lab's format
delivery receipt the buyer receives it
Nothing is delivered until the seller has approved the buyer and the deal is quoted on the ledger. The package is then prepared in the lab's own format by a named adapter, with a delivery receipt carrying the hashes of the manifest and the adapter report. Records served through the API are written to a hash-chained metering log with each content hash.
Check it: Check the hashes on the delivery receipt against the files you received; your usage page is the metering log, aggregated.
-
Payment against a transparent statement, deletion at close-out
statement, deletion certificate the seller receives it
After the buyer pays, the seller statement shows the gross price, the itemised fee and what it covered, any pass-through cost you approved, any tax withheld, and the net. It is read back from the ledger and carries a running total. When the engagement closes, the deposit and its working copies are deleted, the engagement keys are destroyed, and a deletion certificate signed by two people names each store deleted and when, with the access-log export beside it.
Check it: Check the arithmetic, the package hash the statement names, the grant and the running total against your earlier statements.
The rate is the one in your agreement and is itemised on every statement; this site publishes no figure until that rate is set.
What we never do
- Raw deposit bytes never reach a buyer, a web page or an email.
- No statistic below the minimum count is published; no record text reaches a buyer before a signed licence, except a preview watermarked to that buyer.
- Nothing here states that any output is free of personal data or that a legal standard is met. The certificate lists its non-claims.
- Every figure you see is a measurement from your own engagement that you can recompute from the artefacts.