Diagnostic & Shot Embeddings

Embedding shots and diagnostic windows into vectors lets the fabric find physically similar situations by nearest-neighbor search.

Turning shots into vectors

An embedding maps a shot, or a window of diagnostic telemetry, into a fixed-length vector such that physically similar situations land close together. This lets the fabric answer 'what past shots looked like this one?' by nearest-neighbor search — invaluable when a copilot or operator faces an unfamiliar state on the breeder or the burner.

What the embedding captures

• The shape of key features over a window: mode activity, pressure profile, strain deltas.
• Regime membership: ELM-free negative-triangularity for the breeder, plug-and-potential state for the burner.
• Precursor signatures, so shots that preceded a disruption cluster together.
• Machine and configuration context, so retrieval respects which machine it is.

How the vectors are made

Embeddings are produced by models trained on the archive to preserve physical similarity — a shot with a locking mode embeds near other locking-mode shots. Because they are built from the same AI-ready features, they inherit their quality tags and lineage. The vectors are stored in the vector database for retrieval.

Uses

Nearest-neighbor retrieval over shot embeddings supports the anomaly workflow (has this precursor pattern ended badly before?), the copilots (pull similar past operations into context), and MLOps (find drift by watching new shots drift away from known clusters, see quality). The embedding approach is machine-agnostic and is a design specification for the stack serving machines not yet built.