Verified human data. A record on every item, not a badge on the box.
Anyone can print the word verified on a dataset card. This page is what the word means when eeko uses it: nine layers of checks, records, and signatures, built as each item moves from intake through review to delivery, that travel with the data and can be checked without trusting us.
What is verified human data?
Verified human data is training data where three claims are backed by evidence rather than assertion: that a specific person or organisation made the item, that they agreed in writing to its use for AI, and that the item is what it was declared to be, human-made rather than generated, original rather than copied, reviewed rather than merely uploaded. The evidence is attached to the item itself, not kept in a spreadsheet somewhere, and it stays attached when the item is delivered.
The need is not theoretical. Training data is now examined by regulators, by opposing counsel in discovery, and by customers running their own diligence. A vendor's assurance that the data is "clean" answers none of them. A per-item record that a third party can check does.
eeko applies the same record to everything the data practice ships: custom datasets, annotation of your own data, recordings from capture projects, and licensed catalogue work. It is also what makes the practice safe to build on; the agents and retrieval systems we deploy into regulated companies are trained and evaluated on data that can survive the same compliance review as the systems themselves.
Training data provenance, layer by layer.
What happens to an item between the moment it arrives and the moment it ships, and what each step lets you prove afterward.
- SHA-256 and perceptual hashes
- Frame hashes, acoustic fingerprints
- Exact and near-duplicate matching
- Held pending review, resolution logged
- EXIF/XMP, IPTC, C2PA detection
- Detected origin beats declared origin
- Releases and training permission per item
- Licence scope, term, exclusivity
- Block-lists and safety classifier
- Personal data masked, masking recorded
- Label source: expert, owner, or model
- Second-expert pass on an agreed share
- Ed25519 signature, published key
- Hash-chained audit trail, C2PA manifest
- OpenTimestamps proof per original
- Verifiable with any public client
- Forensic mark per delivery
- Receipt, item by item, hash by hash
A provenance record being built.
An illustrative record for one photograph entering the catalogue. Each line is an entry in the item's signed audit trail.
When the question comes, an answer or a document.
A regulator, a litigator, or a customer asks where the training data came from. Here is what each kind of data lets you say.
The vendor said it was fine. Nobody can name the source of an item, show the permission for it, prove a human made it, or show when it existed. The delivery was a folder.
- Source Unknown
- Permission Assumed
- Human-made Unchecked
- Duplicates Likely, unresolved
- Date Unprovable
- Delivery A folder
For any item: the named, consenting rights holder; the signed licence for this use; the checked origin; the fingerprint comparison; a timestamp anchored outside eeko; and the receipt that lists the item by hash.
- Source Named, consenting rights holder
- Permission Signed licence, per item
- Human-made Declared, checked, recorded
- Duplicates Fingerprinted and resolved
- Date Independently anchored
- Delivery Signed receipt, item by item
Four things you can verify without calling us.
A record that only eeko can read is not much of a record. Every layer that matters can be checked with public tools and the files in your delivery.
Declared, signed, anchored, and honest about limits.
How the registry behaves
- Declared — each item's answer on AI training is published where machines look for it: metadata, response headers, robots rules, and a text-and-data-mining reservation. Nobody has to guess.
- Signed — every record carries the registry key's signature, so a changed byte in any entry is detectable by anyone with the public key.
- Anchored — timestamps live in a public blockchain that eeko does not control and cannot edit.
- Honest — a registration is evidence; it is not a legal ruling. Our licences and terms say what the technology proves and what it does not, in plain language.
The same commitments apply whether the item came from a creator licensing their work, a company selling its archive, or a worker in a capture program.
The same proof, whatever the file is.
Fingerprints are chosen to fit the format, so a copy is caught whether it was cropped, re-encoded, trimmed, or transcribed.
Robot episodes are hashed per stream with the operator agreement bound to them; 3D files get shape prints that catch re-exports; text gets prints that find near-copies, with author, reviewer, and rubric version on every demonstration. See teleoperation data and 3D and CAD datasets.
Verification questions.
What does verified mean for training data?
At eeko it means nine specific things happened to the item and each left a record: it was fingerprinted, checked against everything already registered, checked for human origin, bound to its consent and licence terms, screened for safety and for personal data, reviewed by a verified expert, signed, independently timestamped, and listed on a signed receipt. Verified is not a label we attach; it is the sum of those records, and you can inspect every one of them.
How do you detect AI-generated content in a dataset?
Every item declares its making: by a person, with AI assistance, or by a generator. We then look for evidence that contradicts the declaration. Generator signatures in EXIF and XMP metadata, IPTC digital-source-type fields, and embedded C2PA generation records are read automatically, and any of them overrides a human-made claim. The detection is logged on the item, so a buyer can see what was found. Synthetic content is not forbidden, but it is never allowed to pass as human.
What is a signed provenance record?
A structured statement about one item, covering its hashes, its declared and detected origin, its consent and licence references, who reviewed it, and when each of those facts was recorded, signed with the eeko registry key using Ed25519. Every later change to the item is appended to a hash-chained audit trail that is signed the same way, so tampering with any entry breaks the chain. Where the file format supports it, a C2PA manifest carrying the same facts is embedded in the file.
Why anchor timestamps in a public blockchain?
Because a timestamp we issue ourselves only proves what we say. Through OpenTimestamps, the SHA-256 of each original is committed into the Bitcoin blockchain, which nobody, including eeko, can rewrite. Anyone holding the file and the proof can confirm that the file existed by that date without asking us or trusting our systems. For disputes about who made something first, or whether a file predates a model, that independence is the point.
Is a verification record legal proof of ownership?
No, and we say so plainly. Registration with eeko is evidence: it shows what a file looked like, who declared what about it, and when. It is not a copyright ruling and it does not settle disputes by itself. What it gives you is a documented chain you can put in front of counsel, a regulator, or an opposing party, in place of an assurance from a vendor. Our licences and terms are written to be understood, including what the technology cannot do.
What happens when a duplicate is found?
Each upload is matched by exact hash and by perceptual fingerprint against everything already registered, so a crop, a re-encode, or a trimmed clip of an earlier file is still caught. When a match is found, both parties are told, the newer item is flagged, and it stays out of every dataset until a reviewer resolves the conflict. The resolution is written into both items' audit trails, so a buyer can see that it happened and how.
Can we audit the verification ourselves?
Yes. Recompute the SHA-256 of each file and compare it with the receipt. Verify the Ed25519 signature on the provenance record and the audit trail with the published registry key. Check each timestamp proof against the blockchain with any OpenTimestamps client. Open the C2PA manifest with a standard content-credentials tool. None of those steps need eeko to be online or cooperative, which is what makes the record worth having.
Explore the data practice.
Data you can defend.
Tell us what the model needs. Everything we deliver carries the nine layers above, and a data lead replies inside one business day to walk you through the record on a real sample. No deck. A sample and its audit trail.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: