Feedback and Retraining

How pink feedback closes the learning loop: anomalies and drift trigger batch retraining at L0, and validated models return to the twin.

The loop that keeps the twin honest

An AI-native plant is only as good as its ability to correct itself. The feedback-and-retraining loop is how the architecture keeps the twin faithful to a machine that changes — through commissioning, through wear, and across FOAK, NOAK, and BOAK units. It runs entirely above the real-time boundary so it can never disturb control.

The loop

• Detect — L3 anomaly ensembles and drift monitors flag deviations as pink signals.
• Capture — the relevant telemetry, with lineage, is archived at L2 and human corrections are recorded.
• Retrain — batch lines carry the data to L0, which retrains surrogates, GNNs, PINNs, and detectors.
• Validate and promote — MLOps checks the new model against held-out physics before it is allowed near control.
• Deploy — the promoted model updates the twin; the loop continues.

Drift is expected

The twin is trained on simulation before FOAK, so its first contact with real plasma will reveal drift. That is not a failure; it is the loop's purpose. Pink feedback measures the gap between predicted and observed state, and retraining narrows it. Reinforcement learning from human corrections adds operator judgment where models are weak.

Bounded promotion

No retrained model reaches control automatically. Promotion is gated by validation and by the rules engine at L4, so a model that improves average accuracy but violates a safety bound is rejected. The loop makes the plant smarter over time without ever relaxing the guarantees that keep it safe.

The loop's return leg is offline batch lines; its trigger is pink feedback; its substrate is L0.