AI-Ready Structured Features
The fabric's output is a pristine, structured, quality-tagged feature set — the clean interface every L3 model and copilot is built on.
The clean interface
Everything the AI layer does rests on the features the fabric produces. 'AI-ready' means each feature is validated, normalized into physics coordinates, defined by a versioned transform, tagged with a data-quality score, and carrying full lineage. The models never touch raw sensor voltages; they consume this pristine, structured interface.
What makes a feature AI-ready
- Physically meaningful and consistently defined across shots and machines.
- Normalized so a value means the same thing regardless of small equilibrium shifts.
- Quality-tagged, so a model can down-weight imputed or noisy inputs.
- Versioned and lineage-linked, so training and operation stay reproducible.
- Available with online/offline parity from the feature store.
Feeding the twin and copilots
The KRONOS-CTRL twin's modules consume features to run their 50-100 ms predictive shadow; the GNNs consume the sensor-topology features to impute dropped channels; the PINNs consume magnetics and profiles to solve equilibria; the anomaly ensembles consume precursor features; the MPC agents consume the state estimate to plan safe actuation. All of it reads the same clean feature set.
Structure enables retrieval
Because features are structured and consistently described, they can also be embedded and indexed for retrieval (see diagnostic embeddings), letting a copilot pull the most similar past shots into context. Pristine structure is what makes both the numerical models and the retrieval layer possible. The features serve the breeder and the burner alike and are a design specification for machines whose FOAK is expected near 2030.