Loading
A cooperative that trains an open, provenance-emitting model on compute and data its members contribute. Members share in what they build, measured by verifiable contribution.
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.
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.
Rewards settle in dollars through the Citrate compute marketplace. Members share in the long-term value of the model in proportion to verifiable contribution. 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.
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.
Verification runs at the Citrate authorization spine. It is the door, not the surveillance.
The Federation pays real money for real contribution. 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 Sybil accounts and the reward for honest work is not diluted by fake ones.
Verification confirms who a member is. It does not watch what they compute. The architecture keeps those two things separate on purpose.
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.