Coordinating Twin & Copilot

Orchestration reconciles L3 twin predictions and L5 copilot proposals into a single bounded action, resolving disagreement conservatively.

Two intelligences, one machine

The digital twin (L3) predicts machine state and consequences; the copilot (L5) proposes actions to achieve goals. Both are learned, both can be wrong, and they can disagree. Layer 4 is the arbiter: it consumes twin predictions and copilot proposals as events and produces at most one bounded command per actuator per step, or none.

Reconciliation policy

def reconcile(proposal, twin_pred):
    if proposal.confidence < CONF_MIN:      return route_to_human(proposal)
    if twin_pred.uncertainty > UNC_MAX:     return HOLD   # act only on trusted state
    if disagree(proposal, twin_pred):       return conservative_of(proposal, twin_pred)
    return proposal.command                 # still must pass rules + envelope

Conservatism on disagreement

When the copilot's expected outcome conflicts with the twin's prediction beyond a tolerance, orchestration chooses the more conservative action or holds. For the breeder, that biases toward staying inside the current operating point rather than pursuing a proposed higher-performance point on contested predictions. For the burner, it biases toward maintaining established plug confinement over an aggressive efficiency move.

Confidence and uncertainty are inputs, not authority

• Low copilot confidence routes the proposal to a human gate.
• High twin uncertainty holds action until state is re-established with acceptable confidence.
• Neither a high confidence score nor a low uncertainty grants bypass of the envelope check.

The gate is downstream

Reconciliation only selects a candidate command; it does not authorize it. The selected candidate still traverses the full gating pipeline: schema, rules, envelope, approval. This ordering means even a perfectly reconciled, high-confidence proposal is rejected if it would leave the safe region. Both machines remain design-and-simulation studies, so this coordination is exercised against replayed and simulated scenarios today.