Fleet Model Propagation Across Units

As breeder units progress FOAK to NOAK to BOAK, models and lessons propagate across the fleet through the registry, with per-unit validation before any shared model acts.

Many units, one learning system

The breeder program is a fleet: a first-of-a-kind unit followed by next-of-a-kind and best-of-a-kind units. A lesson learned on one unit should benefit all, but no two units are identical, so a model validated on one cannot be assumed valid on another. Fleet propagation is the MLOps process that shares models and data across units while re-validating per unit before granting authority.

Each unit carries its own entry in the fleet registry recording exactly which model version runs where. A model proven on the FOAK unit is offered to a NOAK unit as a challenger, not as an immediate champion: it enters that unit's shadow and canary chain, validated against that unit's own diagnostics and quirks, before it can act there.

What propagates and how

• Data: pulses from all units enrich the shared L0 archive
• Models: a validated model becomes a per-unit challenger, not an auto-champion
• Lessons: regression tests from one unit's incidents guard all units
• Divergence tracking: no two units silently run different behavior unknowingly
• Per-unit envelopes: shared model, unit-specific validated regime

def propagate(model, from_unit, to_unit):
    assert registry.is_prod(model, from_unit)
    # never auto-promote across units; enter target's own chain
    registry.register_for(model, to_unit, state='SHADOW')
    # validate against to_unit's diagnostics + twin before canary
    return f'{model} entered {to_unit} shadow'

Fleet propagation compounds the value of the continual-learning loop: more units mean more data, more regimes explored, and faster model maturation, while the per-unit re-validation ensures scale never erodes safety. Incidents propagate as regression tests through incident-driven retraining, so the whole fleet inherits every unit's hard lessons.