Architecture and verification#
Rizemind addresses two related tasks: training across separate data environments, and maintaining attributable records around that collaboration.
The collaboration problem#
Consider three organizations with data relevant to the same modelling task. A federated workflow allows each organization to train locally and exchange permitted model updates rather than first combine its raw dataset with the others.
Independent participants also need to understand who submitted an update, which round it belongs to, how contributions are measured, and which records can be checked. Rizemind adds mechanisms around the federated workflow for these tasks.
Three functions#
Data and compute: participants operate local training against their own source datasets.
Federated workflow: Flower and the configured Rizemind components coordinate updates, validation, aggregation and evaluation.
Verification: configured components authenticate updates and record selected training metadata. A ledger record is distinct from the underlying dataset and from the model artifact itself.
What Rizemind adds#
Authentication and signed updates. Cryptographic signatures associate an update with a signing identity and allow the signed material to be checked. See signature authentication and the model-signing reference. A signature establishes a relationship to a key; organizational identity and permission to participate require the deployment’s own identity and access policy.
Contribution measurement. Configured strategies evaluate participant contributions against the selected task, metric and evaluation procedure. A contribution score is a result under that method, not an unconditional measure of a participant’s value. See Strategies. Contribution measurement and any compensation mechanism are separate concerns.
Training provenance. Recorded identifiers and metadata can relate participants, rounds, updates, model artifacts and contribution results. See model identity and records for how to distinguish the resulting artifact from a record referring to it.
Ledger-backed records. Selected events or commitments can be recorded onchain. The Arc guide identifies the deployed environment and the verification steps for that example.
Model identity, provenance and usage#
Model identity specifies which model and version a record refers to. Training provenance describes how a version was produced. Usage concerns which version was used in a recorded activity. These are different questions.
When inspecting a record, distinguish a model used as input to an activity from a model produced as output. Interpret its identifiers and checks according to the fields actually captured by the configured workflow. A model reference alone does not establish a subsequent application use.
Data and visibility boundaries#
Local: raw training datasets, local data access, compute and credentials remain under the participant’s administration.
Exchanged: permitted updates, parameters, evaluation outputs and other protocol information follow the configured federated workflow.
Recorded: identifiers, events, signatures or commitments are exposed only as specified by the selected implementation. Review the exact fields before writing them to a public ledger.
Stored elsewhere: metrics and model artifacts may be sent to a configured tracking or artifact service. For example, the MLflow storage module supports logging models and metrics to a tracking server. Local raw data does not mean all derived artifacts remain local.
Raw training datasets are not required onchain. The implementation and its configuration determine the visibility of updates, evaluation outputs and metadata.
Privacy and verification limits#
Keeping source datasets local reduces the need to transfer them; it does not by itself guarantee that updates or derived models disclose no information. Deployment controls may include access restrictions, transport protection, secure aggregation or differential privacy, where supported and configured. Do not assume a protection is enabled because it is available in the broader Flower ecosystem.
For background, see Flower’s differential privacy explanation and secure aggregation explanation.
A signed update or recorded commitment does not, by itself, prove that a model is accurate, that an inference executed correctly, or that a particular model was used in a production system. Those statements require their own mechanisms and evidence.
In these docs, sovereign data describes retaining control over source datasets. It is not a blanket claim of regulatory compliance, data residency or complete infrastructure independence. Each organization must assess its deployment’s requirements and controls.
Scope#
Rizemind is a federated-learning framework with specialized coordination, authentication, contribution-measurement and verification components. It is not a centralized data marketplace, a proprietary foundation model, or an agent runtime. Blockchain records do not make models or data private by themselves.
Next: Illustrative use cases, Integrating with Flower, or Rizemind on Arc Mainnet.