Telemetry Schema Versioning
Channel and feature schemas evolve as the machines mature; versioning ties every archived datum to the exact definitions in force when it was taken.
Definitions change; history must not
Over the life of the breeder and the burner, channels are added, calibrations improve, and feature definitions are refined. If schemas simply mutated, old data would become uninterpretable and reproducibility would collapse. Schema versioning fixes this: every archived record references the schema version in force at acquisition, and old versions are never deleted.
What is versioned
- Channel schemas: the set of ports, their sensors, positions, and calibrations.
- Feature schemas: the definition and inputs of each engineered feature.
- Storage schemas: the layout of archive, feature-store, and vector records.
- Validation and bound schemas: the per-machine limits used to judge data.
Reading old data correctly
When an old shot is re-reduced, the fabric loads the schema version it was taken under, so the raw counts are interpreted with the right calibration and the features are recomputed with the right definitions — or, deliberately, with a newer definition, with both the old and new versions recorded. This is what makes the archive a durable scientific asset.
Compatibility discipline
Schema changes follow compatibility rules: additive changes are safe; breaking changes create a new major version and a migration is recorded. Model versions pin the feature schema versions they were trained on, tying online/offline parity to explicit versions. Versioning is enforced through the registry and applies to both machines. It is a design and operational discipline for machines under construction from Q2 2027.