Edge, cloud, and federated deployment

Match topology to constraints

No single deployment pattern satisfies every workload. Model breeding is useful partly because descendants can target different hardware, privacy, latency, and jurisdiction niches while sharing contracts and lineage.

Edge-only

Use edge-only deployment when latency, offline operation, data locality, or sovereignty dominates. Keep the model population small enough for storage and memory budgets. Prefer routers that do not require cloud metadata on each request. Updates are signed, staged, and reversible.

Cloud-only

Use cloud-only deployment when centralized accelerators, high utilization, or large models dominate. The main risks are network latency, data concentration, and broad blast radius. Isolation and tenant-aware routing remain necessary.

Hybrid

A hybrid cascade runs a local specialist first, then escalates low-confidence or high-complexity cases to a cloud model. The request contract must define which data may leave the device and how local context is minimized before escalation.

Fog or site gateway

A gateway aggregates local telemetry, validates updates, hosts regional specialists, and buffers intermittent connectivity. It can enforce organization-specific policy while the cloud maintains the global registry and evaluation program.

Federated learning or adaptation

Federated systems keep raw data local and exchange updates, but updates can leak information or carry poisoning. Use secure aggregation where appropriate, client attestation, anomaly detection, update clipping, robust aggregation, and a policy for heterogeneous client architectures.

PROCEDURE federated_round(clients, global_parent, policy)
    selected <- SELECT_ATTESTED_CLIENTS(clients, policy.participation_rules)
    updates <- []

    FOR each client IN selected IN PARALLEL
        local_update <- TRAIN_LOCAL_DESCENDANT(global_parent, client.local_data)
        IF LOCAL_GATES_PASS(local_update)
            APPEND updates, SIGN_AND_CLIP(local_update)
        END IF
    END FOR

    filtered <- REJECT_ANOMALOUS_UPDATES(updates, policy.robustness_rules)
    aggregate <- SECURE_AGGREGATE(filtered)
    candidate <- APPLY_UPDATE(global_parent, aggregate)
    RETURN SEND_TO_CENTRAL_EVALUATION(candidate)
END PROCEDURE

Distribution of authority

Edge nodes may adapt caches or approved local adapters but should not publish global champions. Gateways may validate and aggregate. The cloud control plane may propose releases, but site policy can reject them. Document who can create, sign, approve, distribute, and revoke each artifact.

Operational metrics

Measure local and cloud latency, escalation rate, communication bytes, update staleness, client participation, site divergence, rollback success, and quality by hardware tier. A federated system is not healthy merely because aggregation converges.