Train an American open model, together.
The American Learning Federation is a cooperative that trains NAT, Citrate’s open, provenance-emitting model architecture, on compute and data its members contribute. Members share in what they build. Every contribution is measured on the network by verifiable compute and data quality, and rewards follow the measurement.
Membership is identity-verified so payouts reach real contributors and the training pool stays free of fake accounts. The Federation portal is visible to Federation members inside Citrate Core and is still being built out.
SALT has no cash value until mainnet, Q2 2027.
A cooperative that trains an open model.
NAT is Citrate's open model architecture. The Federation is the group that trains it, and shares in it.
NAT partitions a model into six named zones, emits an on-chain-verifiable trace of which zones fired on every forward pass, and stays compatible with standard GGUF and ONNX tooling. It is memory-safe Rust, formally specified, and public. You can read the source and reproduce the results.
The American Learning Federation is the cooperative that trains it. Members contribute compute and data through the Citrate network. The model they train is open, and the people who train it share in what it becomes.
Measured on-chain. Rewards follow the measurement.
No promises about who did the work. The network records it, and the record settles the reward.
Every contribution carries two measured quantities: the compute a member spent and the quality of the data they trained on. The network scores both and computes a reward weight from them on a deterministic fixed-point path, so two nodes and an on-chain verifier compute the same weight from the same contribution, bit for bit. Compute spent on garbage data scores zero and earns zero.
Contribution is measured on-chain, and members share in the long-term value of the model in proportion to it. Rewards settle in SALT, the network's settlement unit, at mainnet, targeted for Q2 2027. The measurement is the entitlement, and the measurement is on the record.
We open this to the public in good faith. The governance and long-term ownership terms are published as the cooperative forms, in plain language, before anyone is asked to commit.
A group can train this architecture in pieces.
Zones are independent. A member can train one zone without retraining the whole model, and the results combine deterministically.
Because each zone owns a fixed slice of the model with a declared contract, a member can train and improve a single zone in isolation. Contributions gather under a signed, verify-before-compose step: every submission is checked before it enters the model, and a malformed or unverified contribution is rejected. The merge is deterministic, so a result computed on one machine reproduces on another and on an on-chain verifier.
In Citrate’s published tests, the partitioned architecture reaches lower loss than a size-matched conventional model at every scale rung measured, and the margin grows with scale, across five seeds per rung. This is an early research result, measured in the open, not a finished proof. The live multi-node training cycle across the network is the current build milestone.
Identity-verified, so honest work isn't diluted.
Verification runs at the Citrate authorization spine. It is the door, not the surveillance.
The Federation rewards real contribution, and a stake in a shared model is worth protecting. That makes it a target for the oldest fraud in the space: one operator spinning up many fake identities to farm rewards for work that was never done. Identity verification closes that door. Every member is a verified, accountable person or institution, so the pool stays free of fake accounts and the reward for honest work is not diluted by duplicate ones.
Verification confirms who a member is. It does not watch what they compute. The architecture keeps those two things separate on purpose.
Contribute compute and data. Share in the model.
Membership is identity-verified and opens at the Citrate authorization spine. The first step is an application. There is no obligation, and the cooperative terms are published before anyone commits.