STEP 01
WHERE IT STAYS
Train locally
Each approved party trains on its own data, inside its own boundary. Records never cross the wall.
→AI LABS · FEDERATED LEARNING · SHARE THE UPDATES, NOT THE DATA
Learn together. Keep it separate.
Share the updates. Not the data. Train across parties without moving records out of the rooms they belong in, and keep the proof on the round.
PARTIES
Approved first
Every collaborator visible before the round can start.
UPDATES
Not records
Local sites train. Only the model updates travel.
PROOF
On the round
Aggregation proof and budget left, on one record.
HOW IT WORKS
Three steps. No raw transfer.
Train locally. Share updates only. Aggregate the gain, with the proof on the round.
STEP 02
WHAT TRAVELS
Share updates only
Updates leave the site, records don't. Secure aggregation and a privacy budget keep the spend-down visible.
→STEP 03
WHAT WE SIGN
Aggregate the gain
The round leaves with a proof hash, participant health, and remaining budget — attached to the release.
A ROUND, READ AS A SIGNAL
Approvals in. Updates back. Proof out.
A round reads like every other release on the platform. Signal, review, gate — with the budget and the proof attached.
WHAT COMES OUT
What every round leaves with.
Federation reads like the rest of the platform — every round leaves a record the next one has to clear.
01
Update payloads
Local updates submitted by each approved party, never the records behind them.
↳ ARTIFACT
02
Privacy proofs
Differential privacy budget spend-down and remaining headroom, on every round.
↳ ARTIFACT
03
Aggregation logs
Who joined, who dropped, communication cost, and the secure aggregation status.
↳ ARTIFACT
04
Signed rounds
Proof hash, participant health, and verdict — attached to the release record.
↳ ARTIFACT
WHERE IT FITS
In the loop, this is how you remember.
Test the run. Review the hard cases. Recruit the right specialist. Remember what each party can share. Approve what's right.
01
Test
02
Review
03
Recruit
04
Remember
● YOU ARE HERE
05
Approve
RELATED MODULES
Next to this in the Model OS.
TRAINING
Train on what your team trusts.
Fine-tune with the rubric, the reviewers, and the data you already keep.
See the page →RL ENVIRONMENTS
Practice in a room that remembers.
Deterministic environments for agents that need to be tested before they ship.
See the page →SYNTHETIC POPULATIONS
Stand in for the rare case.
Generated populations that fill the gaps without crossing the wall.
See the page →FEDERATED LEARNING
Learn together. Keep it separate.
Bring the parties. Bring the boundaries. We'll handle the rounds, the budget, and the proof.