Fleet Learning and Model Propagation

Validated models learned on one unit propagate to others under governance, so each new breeder or burner starts from the accumulated experience of the fleet.

Learning once, applying everywhere

The strategic value of a fleet is that experience compounds. A degradation signature, a control refinement, or a physics correction learned on one unit should benefit every comparable unit. L7 fleet learning collects validated model updates from units and propagates them under governance to the rest of the fleet, so a NOAK unit begins with everything FOAK learned and a BOAK unit with everything before it.

Governed, not automatic

Propagation is gated, not blind. A model update learned on one unit is validated against the receiving unit's own twin and history before it is allowed to influence control, because units differ by generation, configuration, and wear. Updates that improve a shared physics surrogate (equilibrium, neutronics, RUL) propagate readily; updates tied to a specific unit's idiosyncrasy do not.

# federated-style propagation with per-unit validation gate
def propagate(update, source_unit):
    for u in fleet_peers(source_unit):
        if not compatible(update, generation(u), config(u)):
            continue
        score = validate(update, twin(u), holdout_history(u))
        if score >= accept_threshold and safety_review(update, u):
            stage(update, u)          # into shadow, then L3 after soak
# raw telemetry stays local; only validated model deltas move

Raw plant telemetry does not need to move between units for learning to happen; validated model deltas do. This federated pattern keeps each unit's high-rate data local (respecting security boundaries, especially for Aegis defense installations) while still letting the fleet share what was learned from it.

Propagation direction follows the generation gradient but is not one-way. Early FOAK experience seeds NOAK and BOAK; but a later, better-instrumented unit can also correct a model the earlier generation carried. L7 records the provenance of every propagated update so a fleet-wide model change is traceable to the unit and data that justified it.

Fleet learning runs today across twin-modeled units and historical device data, structured so that real early-unit experience from FOAK ~2030 immediately improves the models later units inherit. No performance is claimed for a later unit beyond what validated propagation supports.