Technology: The Goose FL Trust Stack

Institutions in finance, healthcare, energy and government hold the most valuable data in the economy, and almost none of it can move. Privacy law, commercial secrecy and institutional firewalls see to that. The Goose FL trust stack is federated learning infrastructure that trains shared models across institutions while the data stays exactly where it is, wrapped in the guarantees that regulated, multi-institution collaboration actually requires.

Open FL core

Federated orchestration assembled from proven open-source standards, pinned and security-patched. The model travels to each institution; raw data never does.

Assembled · open standards

2
Trust layers
The layers Goose FL builds: personalization for real-world data, privacy that does not punish small institutions, erasure executed inside trained models, and oversight regulators can inspect.
Proprietary · Goose FL engineering

3
Agentic governance
The NOMOS architecture: a checkpoint between AI agents and the world, so every automated action is approved, logged and auditable.
Proprietary · NOMOS architecture

Layer 1: the open core

Built on Flower, the open-source standard for federated learning (version-pinned), with TensorFlow for local model training. Flower provides the orchestration: a coordination server and one training client per participating institution. Goose FL contributes the deployment discipline around it.

Secure aggregation: model updates are combined under secure multiparty computation, so no participant, and not even the operator, can inspect any single institution's contribution.
Edge nodes: each institution runs a local node that connects to its existing systems. In healthcare that means FHIR R4 integration and OMOP CDM standardization; in finance, containerized clients behind the institution's own firewall. Local inference returns results in under 200 ms, inside the EHR or the institution's own systems.
Sovereign deployment: fully containerized services with GitOps delivery, deployable on national infrastructure under local jurisdiction. Data residency is a property of the architecture, not a promise in a contract.

Verified in deployment: the secure aggregation workflow has been exercised end to end in a live multi-node configuration: three supernodes, twenty federated rounds, zero failures (August 2026).

Layer 2: the trust layers

Open-source software gets the federated network up and running. But to make it secure enough for regulated institutions, Goose FL built four proprietary layers of security and compliance on top.

pFedMe personalization

Real-world data is never uniform: every hospital, bank and agency has a different population. Standard averaging produces one model that fits none of them. Each institution receives a model personalized to its own community while still learning from the group. Validated in cancer-detection experiments on real clinical data. See the results.

Privacy equity architecture

Standard differential privacy applies one noise level to everyone, which degrades models precisely where data is scarcest. Ours is different: decoupled noise applied at aggregation rather than inside each trainer, per-client calibration so smaller centers get proportionally less noise, and gradient sparsification so the same guarantee costs less noise.

Machine unlearning

A data subject's erasure right is honored inside the trained model itself: certified retraining with a completion certificate per request. The model the institution actually serves is the model the erasure applies to.

Accountable oversight

Privacy budgets tracked per round and per window; tamper-evident, hash-chained audit logs for every training decision; and supervisory dashboards that show regulators model performance and privacy consumption without ever exposing raw data.

Layer 3: agentic governance

Institutions are beginning to run AI agents alongside their data infrastructure. Agents act fast, act autonomously, and fail differently from conventional software, so the stack governs them with the NOMOS architecture: a checkpoint that sits between an AI agent and the world.

Agent registers→ Action submitted before execution→ Policy engine: allow / block / escalate→ Decision written to tamper-evident log

Every decision lands in a SHA-256-chained audit log. Per-agent trust scoring with decay and anomaly detection, quarantine, circuit breakers, and a system-wide kill switch contain an agent that misbehaves. Compliance reporting is generated from live control state, aligned to the EU AI Act, ISO/IEC 42001 and the NIST AI Risk Management Framework. Agents integrate through the API or MCP.

Every automated action is approved, logged, and fully auditable. See how the checkpoint works.

Explore the NOMOS governance architecture →

A federated round, in four steps

Step 1

Distribute

The shared model travels out to every participating institution. No data moves with it.

Step 2

Train locally

Each institution trains behind its own walls, on its own customers or patients.

Step 3

Protect and combine

Updates are privacy-noised and combined under secure computation. No individual update is inspectable.

Step 4

Return

An improved model flows back, personalized for each institution's own population.

Want the interactive version? Federated Learning 101 walks the same cycle with a live simulation and a self-test.

Design principles

The model travels. The data never does. There is no central vault of anyone's data, so there is no honeypot to attack.
Math replaces trust. Institutions collaborate through cryptographic guarantees, not through promises.
Protection that does not punish the small. Privacy is tuned per institution, because blanket noise hurts the communities with the least data the most.
Governance all the way through. Every training round, every agent action and every erasure request leaves an auditable record.
People decide. The infrastructure recommends; institutions and regulators decide.

Learn more