Datasets belong to the network.

I. The thesis

Frontier AI runs on data labeled by people. Every model that matters at scale was trained on a dataset that someone, somewhere, manually verified. That labor is the actual infrastructure of modern AI. We have decided it deserves the structure of infrastructure — addressable, signed, paid, owned.

The Network exists because the alternative — invisible labor, opaque vendors, datasets that rot inside private buckets — is no longer the only available shape. There is another shape. We are building it.

II. The labor problem

The current vendor model has three quiet failures. The first is opacity. A buyer ships data to a vendor and receives labels back. What happened in between — who labeled, how it was checked, who got paid, what fraction of the buyer's spend reached the person who did the work — is not part of the contract. The buyer trusts. The vendor answers. Quality is a status report, not a verifiable property of the artifact.

The second is gig-work economics. The people who actually label data — Humans, in our lexicon, since they are the only Humans the system has — are routed through layers of subcontracting that compress hourly rates and eliminate any sense of contribution. The work is interchangeable, throwaway, anonymous. The pay reflects that framing.

The third is the vanishing dataset. Once a model is trained, the labeled corpus usually disappears into a private database. It cannot be re-audited. It cannot be reused. It cannot be reproduced. The labor that produced it leaves no trace. A field that takes data seriously cannot afford a foundation that vanishes the moment it is useful. We treat each of these three failures as a design constraint, not a tradeoff to be managed. The vendor model has settled into them; the Network refuses them as starting conditions.

III. The Humaniti inversion

We invert all three. Labor is paid directly, per verified label, with the rate visible on the page before any work begins. Verification is consensus — N independent Humans label the same item, an aggregator weighs each vote by per-skill Elo and prior accuracy, a Steward adjudicates disagreement, and gold questions catch drift on every batch. The dataset is signed and addressable — every label carries a receipt of the Humans who produced and verified it, and the audit chain is sha256-linked so that tampering is detectable on a single scan.

None of this is exotic. The components are well-understood. Per-skill Elo has decades of literature behind it. Dawid-Skene aggregation has been around since the seventies. Append-only audit chains are how every serious financial system has worked for years. What is new is the commitment to ship them as the default rather than as a premium feature. The Network does not have a tier where verification is optional. There is one tier. Quality is a protocol property, not a vendor promise.

The same inversion applies to the people who do the work. A Human is not an anonymous vendor account. Each Human carries a verifiable accuracy history, an Elo per skill, a tier, and a payout record. Verified accuracy translates directly into higher per-label rates and access to higher-stakes work. The labor accumulates into a portable record of competence — and the per-label receipts make that record auditable by anyone, not only by us.

IV. The arc

Phase 1 is the Network — verified Humans, verified work, direct pay. That is what ships first. It is the substrate everything else builds on. Phase 2 is the Record — datasets become signed, addressable, permanent records that survive the company that produced them. Phase 3 is the Economy — Humans accumulate credits in the work they verify, stake them, multiply their pay rate, and qualify for Steward roles. Phase 4 is the Commons — Humans and Builders govern the Network together; Humaniti the company becomes a steward, not an owner.

The through-line is the same in every phase. Labor that produces value should accumulate ownership in what it produces, and governance over the rules of its own labor. The four phases are the same idea, repeated at four levels of depth. Read the full arc →

V. The invitation

If you build models, you already pay for labels. The question is whether you pay a vendor for an opaque process or a network for a verifiable one. If you label, you already do this work. The question is whether you are paid by a subcontractor on net-60 terms or by a network on per-verified-label terms with your name on the output. Either way, the work is the same. The shape of the work is what we are changing.

We are not asking anyone to take this on faith. The audit chain is a current shipped feature. The consensus pipeline is a current shipped feature. The direct-pay rails are a current shipped feature. The pieces of Phase 2, 3, and 4 land on a public roadmap with public dates. If we miss a date, the date moves and we say so.