Document & Schema Registry

A registry holds the authoritative definitions of every channel, feature, and document, so the whole fabric agrees on what each number means.

The single source of definitions

A fabric with 60+ channels, dozens of engineered features, and two machines needs one authoritative place that says what each thing is: the channel's sensor, calibration, and units; the feature's definition and version; the document's provenance. The registry is that place, and every stage of the fabric reads from it so definitions never fork.

What the registry holds

• Channel definitions: sensor type, position, feedthrough, calibration, sample rate, filter cutoffs.
• Feature definitions: the exact transform, its inputs, and its version.
• Schema versions for archive, feature store, and vector database records.
• Document provenance for the corpus the copilots retrieve over.
• Per-machine bounds and envelopes for validation and safety.

Versioned, not mutable

Definitions change as the machines and models mature, so the registry is versioned rather than mutable. An archived shot references the schema versions in force when it was taken; a model references the feature versions it trained on. This is what makes re-reduction and reproducibility possible across the life of both the breeder and the burner.

Grounding the copilots

The registry is also part of the retrieval corpus: a copilot grounds its answers in current channel and feature definitions by retrieving from the registry, so it never reasons over a stale schema. Together with metadata governance and lineage, the registry is the backbone of the fabric's trustworthiness. It is a design specification for the stack of machines under construction from Q2 2027.