Why Now
The case for putting data ownership and attribution on-chain: what the current labelling market cannot do, and which of those gaps the protocol actually closes.
Labelling vendors centralise both the data and the profit from it. A contributor is paid once, holds no ownership of what they produced, gets no durable credit for it, and has little visibility into how their work was checked. That suppresses quality, makes rigorous validation an expense rather than an asset, and fails to attract the people whose judgment is worth the most.
The sharpest version of the problem is that these pipelines are not built for expert contributors. Finding and keeping lawyers, clinicians and researchers requires an incentive that matches the effort — and a one-off fee with no attribution does not.
Why demand is spiking now
- Expert data is the bottleneck. Enterprises need vertical-grade data for frontier models and cannot reliably mobilise or retain high-skill annotators to produce it.
- Verifiable annotation is infrastructure. Quality claims about training data are now something buyers and regulators ask to see evidence for, not something a vendor asserts.
- The economics fit. Data has durable value realised over time through licensing and access, which makes a royalty a better-shaped instrument for it than a purchase price.
What the protocol changes
Ownership instead of a one-off fee. A contributed, validated piece of work becomes part of a dataset version whose ownership is divisible. Holding a fraction entitles you to a share of everything paid in for that dataset afterwards — continuously, and without having to be paid again for the original work. That is the direct fix for "paid once".
Attribution that resolves. Each contribution is anchored as a fingerprint naming its contributor, its content hash and the task it was done for, and the contract rejects a task that does not exist. So contribution → task → campaign → frontier is a chain anyone can walk, not a provenance claim in a document. Credit is a property of the record rather than a courtesy.
Earnings that survive the things that usually break them. Claiming your shares late loses you nothing: a distribution is divided by the share total fixed at assembly, so unclaimed shares are reserved rather than shared out among whoever claimed first. Selling a fraction does not hand over the revenue that accrued while you held it. Rotating the wallet behind your identity redirects future claims with no migration.
Validation is recorded, not asserted. Anchoring a contribution requires a validator's signature over its content, and the verdict and quality grade are written onto the record. The assessment is part of the artefact rather than a report alongside it — how much signal a given deployment's grades carry is a separate question, and one about that deployment's validators rather than about the protocol.
Access is revocable at the source. A grant of access is anchored on-chain and can be revoked there, whatever credentials are already in circulation — which is what makes licensing sensitive data a different proposition from sending someone a copy of it.
Who contributes, and what they get
| Contributor | What blocks them today | What the protocol offers |
|---|---|---|
| Domain experts (MD / JD / PhD) | One-off fees, no attribution, high opportunity cost | A fraction of every dataset their work is assembled into, and attribution that resolves on-chain |
| Curators and assemblers | No stake in the datasets they shape | No frontier owner gates assembly, and the assembler sets the share allocation the version commits to |
| Validators | Quality work is a cost centre, and its results are invisible | The verdict and grade they sign are written onto the record, and a validation is itself a contribution that can carry weight in a version's allocation |
| Platforms (labelling, RAG, tooling) | No share in downstream value, and re-architecting to get one is prohibitive | Integrate against the contracts and SDKs and keep your own product surface |
Legacy versus this protocol
| Dimension | Labelling vendor | XnY Protocol |
|---|---|---|
| Contributor upside | One-off payment | A fraction of revenue paid in for the dataset, claimable indefinitely |
| Attribution | Centralised owner, opaque provenance | Fingerprints resolving to task, campaign and frontier on-chain |
| Validation | Sampled internally, results not published | Validator signature, verdict and grade on the record itself |
| Access | A copy is delivered | A grant, anchored on-chain and revocable at the source |
| History | Corrections overwrite | Append-only: a correction is a new record pointing at the old one |
What this page no longer claims
Earlier versions of this page rested on mechanisms the protocol does not have: Train-Now-Pay-Later licensing, staking-as-confidence on data and on contributors, slashing of those stakes, blinded peer review with automatic escalation, cross-attestation panels, post-deployment challenge windows that reallocate royalties, a reputation score, and a Long-Term Success Index driving task routing and revenue splits.
None of it is implemented, and most of it has no design behind it either. The case above is deliberately narrower, and made only from what the contracts do. See Future Directions for the aspirations that are being kept, and kept clearly marked.
Last updated