The Low-Latency Feature Store

A feature store serves engineered features to control, the twin, and the copilots at low latency, guaranteeing the same values online and offline.

One source of features

The feature store is the fabric's serving layer: the single place that engineered features — instantaneous Q, core pressure map, strain deltas, plug density — are read from, whether by the low-latency online path feeding control or the offline batch path building training sets. Making it the one source is what guarantees a model sees the same feature in training and in operation.

Online and offline

• Online: features computed and served fast enough to feed the twin's predictive shadow and the sub-10 microsecond control boundary.
• Offline: the same feature definitions applied to archived series for retraining.
• Parity is enforced by construction (see online/offline parity).
• Every feature is served with its data-quality score and lineage handle.

Latency budget

The store must serve within the loop's timing budget. It keeps recent features in memory, indexed by machine, channel set, and time, and precomputes the reductions that the twin requests repeatedly. Features that require heavy inversion (the pressure map) are computed once and shared, not recomputed per consumer.

Consumers

The twin's Power, Neutronics, Thermomechanics, and MHD modules read from the store; the anomaly ensembles read precursor features from it; the copilots read operational features for dispatch and scheduling. Because it is the shared read path for both the breeder and the burner, the store enforces the schema and quality contracts uniformly. Its design is validated in simulation for machines whose FOAK operation is expected near 2030.